llms-beyond-chat.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "# LLMs: Beyond Chat\n",
        "\n",
        "Large Language Models, like GPT-4o, are well known as capabale chatbots, through applications like ChatGPT and Copilot. While chatting with a computer is a magical experience, these models can do some much more than just chat. In this tutorial we'll review some of these capabilities and adopt an approach for working with LLMs that treats them more like an execution engine for software - a VM, if you will - than a chatty persona. We'll use structured input and ouput, relying on typed schema, to interface between the textual and difficult to predict and control world of LLMs and the realm of software development."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "aI-zWHmskQru"
      },
      "source": [
        "## Setting Up"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "iRxzoZEYkQrv"
      },
      "source": [
        "### Installing Dependencies\n",
        "\n",
        "We'll need the following Python packages:\n",
        "- `openai` for making calls to Azure Open AI\n",
        "- `pydantic` and `instructor` for using typed, structured input and output with our LLM calls\n",
        "- `pydantic-settings` for loading the endpoint configuration from the .env file\n",
        "- `pandas`, `matplotlib`, and `networkx` to help us visualise our output"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "id": "N5hLfrJGkQrw"
      },
      "outputs": [],
      "source": [
        "%pip install openai pydantic pydantic-settings instructor pandas matplotlib networkx\n",
        "from IPython.display import clear_output ; clear_output()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "MVV6MCRIkQrw"
      },
      "source": [
        "### Configuration\n",
        "\n",
        "Copy the file `dot.env` to `.env`. Edit it and update the `OPENAI_API_KEY` if you are using OpenAI, or if you are using Azure Open AI, update the values for `AZURE_OPENAI_ENDPOINT`,`AZURE_OPENAI_API_KEY`, and `GPT_4_O_MODEL_NAME` (that's your deployment name).\n",
        "\n",
        "The next cell will load these values and configure the OpenAI SDK to use them."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "d2dCRzUxkQrx"
      },
      "outputs": [],
      "source": [
        "from pydantic_settings import BaseSettings, SettingsConfigDict\n",
        "from pydantic import SecretStr\n",
        "from pydantic import model_validator\n",
        "from pprint import pprint\n",
        "\n",
        "class Settings(BaseSettings):\n",
        "    model_config = SettingsConfigDict(env_file='.env')\n",
        "\n",
        "    OPENAI_API_KEY: SecretStr\n",
        "    AZURE_OPENAI_ENDPOINT: str\n",
        "    AZURE_OPENAI_API_KEY: SecretStr\n",
        "    GPT_4_O_MODEL_NAME: str = 'gpt-4o'\n",
        "\n",
        "    @model_validator(mode='after')\n",
        "    def validate_api_config(cls, values):\n",
        "        if len(values.OPENAI_API_KEY) == 0 and len(values.AZURE_OPENAI_API_KEY) == 0:\n",
        "            raise ValueError('API Key (either OPENAI_API_KEY or AZURE_OPENAI_API_KEY) is required')\n",
        "        if len(values.AZURE_OPENAI_API_KEY) > 0:\n",
        "            if len(values.AZURE_OPENAI_ENDPOINT) == 0:\n",
        "                raise ValueError('AZURE_OPENAI_ENDPOINT is required when using Azure Open AI')\n",
        "        if len(values.GPT_4_O_MODEL_NAME) == 0:\n",
        "            values.GPT_4_O_MODEL_NAME = 'gpt-4o'\n",
        "        return values\n",
        "\n",
        "    @property\n",
        "    def is_azure(self):\n",
        "        return (len(self.model_config.get('AZURE_OPENAI_ENDPOINT', '')) > 0 and\n",
        "                len(self.model_config.get('AZURE_OPENAI_API_KEY', '')) > 0)\n",
        "\n",
        "    def get_oai(self):\n",
        "        if self.is_azure:\n",
        "            from openai import AzureOpenAI\n",
        "            return AzureOpenAI(\n",
        "                api_version=\"2024-05-01-preview\",\n",
        "                azure_endpoint=self.AZURE_OPENAI_ENDPOINT,\n",
        "                api_key=self.AZURE_OPENAI_API_KEY.get_secret_value(),\n",
        "            )\n",
        "        else:\n",
        "            from openai import OpenAI\n",
        "            return OpenAI(api_key=self.OPENAI_API_KEY.get_secret_value())\n",
        "\n",
        "settings = Settings()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "t2FAs_fKkQrx"
      },
      "source": [
        "## Using Structured Input/Output with OpenAI\n",
        "\n",
        "To use typed, structured input and output with our LLM calls, we will be relying on `pydantic` and `instructor`.\n",
        "\n",
        "**Pydantic** ( [https://docs.pydantic.dev/](https://docs.pydantic.dev/) ) is a popular package for extending Python's typing system with declarative interfaces.\n",
        "\n",
        "**Instructor** ( [https://python.useinstructor.com/](https://python.useinstructor.com/) ) patches the LLM SDK with the ability to use a Pydantic model for specifying the JSON schema for the LLM output, and parsing the result into the Pydantic mode. We will be using it in all of our examples to interact with our LLM.\n",
        "\n",
        "Let's start by patching our LLM client and defining some helper functions..."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 3,
      "metadata": {},
      "outputs": [],
      "source": [
        "import instructor\n",
        "from pydantic import BaseModel, Field\n",
        "from typing import List\n",
        "from enum import Enum\n",
        "import json\n",
        "import pandas as pd\n",
        "import networkx as nx\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "# `client` is a patched Open AI SDK client that allows passing a Pydantic\n",
        "# model for specifying the schema and parsing the output.\n",
        "client = instructor.from_openai(settings.get_oai())\n",
        "\n",
        "# Let's define a helper function for calling the LLM. We will use this\n",
        "# function for all our LLM calls.\n",
        "def llm(response_model: BaseModel = BaseModel, system: str = None,\n",
        "        user: str = None, temperature: float = 0.0, max_tokens: int = 1000):\n",
        "    \"\"\"\n",
        "    Helper function for calling the LLM (GPT-4o) with a Pydantic BaseModel,\n",
        "    a system prompt and/or a user prompt, with temperature and max_tokens.\n",
        "    \"\"\"\n",
        "    messages = []\n",
        "    if system:\n",
        "        messages.append({\"role\": \"system\", \"content\": system})\n",
        "    if user:\n",
        "        messages.append({\"role\": \"user\", \"content\": user})\n",
        "    result = client.chat.completions.create(\n",
        "        model=settings.GPT_4_O_MODEL_NAME,\n",
        "        response_model=response_model,\n",
        "        messages=messages,\n",
        "        temperature=temperature,\n",
        "        max_tokens=max_tokens,\n",
        "    )\n",
        "    return result\n",
        "\n",
        "def print_schema(model: BaseModel):\n",
        "    \"\"\"\n",
        "    Print the JSON schema corrsponding to a Pydantic model.\n",
        "    \"\"\"\n",
        "    print(json.dumps(model.model_json_schema(), indent=2))\n",
        "\n",
        "def print_result(result: BaseModel):\n",
        "    \"\"\"\n",
        "    Print the Pydantic model result of an LLM call as JSON.\n",
        "    \"\"\"\n",
        "    print(result.model_dump_json(indent=2))\n",
        "\n",
        "# Configure Pandas to format dataframes for pretty output inside the notebook\n",
        "pd.set_option('display.max_columns', None)\n",
        "pd.set_option('display.expand_frame_repr', False)\n",
        "pd.set_option('display.max_colwidth', None)\n",
        "pd.DataFrame._repr_html_ = lambda df: df.style.set_properties(**{'text-align': 'left'})._repr_html_()\n",
        "\n",
        "def visualize_graph(graph):\n",
        "    \"\"\"\n",
        "    Visualize a graph.\n",
        "    Expects a graph object with `nodes` and `edges` properties,\n",
        "    where each node has an `id`, `label`, and `color`\n",
        "    and each edge has `source`, `target`, `label`, and `color`.\n",
        "    \"\"\"\n",
        "    G = nx.DiGraph()\n",
        "    for node in graph.nodes:\n",
        "        G.add_node(node.id, label=node.label, color=node.color)\n",
        "    for edge in graph.edges:\n",
        "        G.add_edge(edge.source, edge.target, label=edge.label, color=edge.color)\n",
        "    pos = nx.planar_layout(G)\n",
        "    node_colors = [node[1]['color'] for node in G.nodes(data=True)]\n",
        "    edge_colors = [edge[2]['color'] for edge in G.edges(data=True)]\n",
        "    labels = {node[0]: node[1]['label'] for node in G.nodes(data=True)}\n",
        "    plt.figure(figsize=(7, 5))\n",
        "    nx.draw(G, pos, labels=labels, with_labels=True, node_color=node_colors, edge_color=edge_colors, font_size=8)\n",
        "    plt.title(\"Knowledge Graph\")\n",
        "    plt.show()\n",
        "\n",
        "def print_tree(data, indent=0):\n",
        "    \"\"\"\n",
        "    Pretty-print a JSON object as a tree.\n",
        "    \"\"\"\n",
        "    if isinstance(data, dict):\n",
        "        for key, value in data.items():\n",
        "            print('   ' * indent + key.upper() + ':')\n",
        "            print_tree(value, indent + 1)\n",
        "    elif isinstance(data, list):\n",
        "        for item in data:\n",
        "            print_tree(item, indent)\n",
        "    else:\n",
        "        print('   ' * indent + str(data))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UQBHip4rkQry"
      },
      "source": [
        "## Working with Text\n",
        "\n",
        "LLMs are master text manipulators. If it's reading or writing text, the best LLMs can do amazing things."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Translation and Normalization\n",
        "\n",
        "Reading and writing, understanding and translating text between languages, or language styles, is an easy task with LLMs.\n",
        "\n",
        "Let's start by reading a sentence in one language, detecting which language it is, and translating the sentence to English."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 4,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "SCHEMA:\n",
            "{\n",
            "  \"properties\": {\n",
            "    \"input_language\": {\n",
            "      \"description\": \"The language of the original text, as 2-letter language code.\",\n",
            "      \"title\": \"Input Language\",\n",
            "      \"type\": \"string\"\n",
            "    },\n",
            "    \"translation\": {\n",
            "      \"title\": \"Translation\",\n",
            "      \"type\": \"string\"\n",
            "    }\n",
            "  },\n",
            "  \"required\": [\n",
            "    \"input_language\",\n",
            "    \"translation\"\n",
            "  ],\n",
            "  \"title\": \"TranslatedString\",\n",
            "  \"type\": \"object\"\n",
            "}\n",
            "RESULT:\n",
            "{\n",
            "  \"input_language\": \"de\",\n",
            "  \"translation\": \"Language skills are an important part of communication.\"\n",
            "}\n"
          ]
        }
      ],
      "source": [
        "german_text = \"Sprachkenntnisse sind ein wichtiger Bestandteil der Kommunikation.\"\n",
        "\n",
        "class TranslatedString(BaseModel):\n",
        "    input_language: str = Field(\n",
        "        ...,\n",
        "        description=\"The language of the original text, as 2-letter language code.\"\n",
        "    )\n",
        "    translation: str\n",
        "\n",
        "print(\"SCHEMA:\")\n",
        "print_schema(TranslatedString)\n",
        "\n",
        "translation = llm(\n",
        "    TranslatedString,\n",
        "    \"Detect the language of the original text and translate it into English.\",\n",
        "    german_text,\n",
        ")\n",
        "\n",
        "print(\"RESULT:\")\n",
        "print_result(translation)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Our LLM can cope with a more complex task. Let's get it to translate the same sentence to multiple languages.\n",
        "\n",
        "Comment on token selection: for many processing tasks we'd want the LLM to be quite conservative with token selection, to get the most accurate results, but when it comes to writing text, it is often better to give it more freedom in selecting tokens from a wider distribution. We can control that by passing a higher `temperature` value when making the call."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 5,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "{\n",
            "  \"$defs\": {\n",
            "    \"TargetLanguage\": {\n",
            "      \"enum\": [\n",
            "        \"de\",\n",
            "        \"fr\",\n",
            "        \"it\",\n",
            "        \"es\",\n",
            "        \"he\"\n",
            "      ],\n",
            "      \"title\": \"TargetLanguage\",\n",
            "      \"type\": \"string\"\n",
            "    },\n",
            "    \"Translation\": {\n",
            "      \"properties\": {\n",
            "        \"language\": {\n",
            "          \"allOf\": [\n",
            "            {\n",
            "              \"$ref\": \"#/$defs/TargetLanguage\"\n",
            "            }\n",
            "          ],\n",
            "          \"description\": \"The language of the translated text, as 2-letter language code.\"\n",
            "        },\n",
            "        \"translation\": {\n",
            "          \"title\": \"Translation\",\n",
            "          \"type\": \"string\"\n",
            "        }\n",
            "      },\n",
            "      \"required\": [\n",
            "        \"language\",\n",
            "        \"translation\"\n",
            "      ],\n",
            "      \"title\": \"Translation\",\n",
            "      \"type\": \"object\"\n",
            "    }\n",
            "  },\n",
            "  \"properties\": {\n",
            "    \"translations\": {\n",
            "      \"items\": {\n",
            "        \"$ref\": \"#/$defs/Translation\"\n",
            "      },\n",
            "      \"title\": \"Translations\",\n",
            "      \"type\": \"array\"\n",
            "    }\n",
            "  },\n",
            "  \"required\": [\n",
            "    \"translations\"\n",
            "  ],\n",
            "  \"title\": \"Translations\",\n",
            "  \"type\": \"object\"\n",
            "}\n"
          ]
        },
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_77b71_row0_col0, #T_77b71_row0_col1, #T_77b71_row1_col0, #T_77b71_row1_col1, #T_77b71_row2_col0, #T_77b71_row2_col1, #T_77b71_row3_col0, #T_77b71_row3_col1, #T_77b71_row4_col0, #T_77b71_row4_col1 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_77b71\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_77b71_level0_col0\" class=\"col_heading level0 col0\" >language</th>\n",
              "      <th id=\"T_77b71_level0_col1\" class=\"col_heading level0 col1\" >translation</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_77b71_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_77b71_row0_col0\" class=\"data row0 col0\" >TargetLanguage.de</td>\n",
              "      <td id=\"T_77b71_row0_col1\" class=\"data row0 col1\" >Große Sprachmodelle sind ein leistungsstarkes Werkzeug für die Verarbeitung natürlicher Sprache.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_77b71_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_77b71_row1_col0\" class=\"data row1 col0\" >TargetLanguage.fr</td>\n",
              "      <td id=\"T_77b71_row1_col1\" class=\"data row1 col1\" >Les modèles de langage larges sont un outil puissant pour le traitement du langage naturel.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_77b71_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_77b71_row2_col0\" class=\"data row2 col0\" >TargetLanguage.it</td>\n",
              "      <td id=\"T_77b71_row2_col1\" class=\"data row2 col1\" >I grandi modelli linguistici sono uno strumento potente per l'elaborazione del linguaggio naturale.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_77b71_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_77b71_row3_col0\" class=\"data row3 col0\" >TargetLanguage.es</td>\n",
              "      <td id=\"T_77b71_row3_col1\" class=\"data row3 col1\" >Los grandes modelos de lenguaje son una herramienta poderosa para el procesamiento del lenguaje natural.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_77b71_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_77b71_row4_col0\" class=\"data row4 col0\" >TargetLanguage.he</td>\n",
              "      <td id=\"T_77b71_row4_col1\" class=\"data row4 col1\" >מודלים לשוניים גדולים הם כלי רב עוצמה לעיבוד שפה טבעית.</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "            language                                                                                               translation\n",
              "0  TargetLanguage.de          Große Sprachmodelle sind ein leistungsstarkes Werkzeug für die Verarbeitung natürlicher Sprache.\n",
              "1  TargetLanguage.fr               Les modèles de langage larges sont un outil puissant pour le traitement du langage naturel.\n",
              "2  TargetLanguage.it       I grandi modelli linguistici sono uno strumento potente per l'elaborazione del linguaggio naturale.\n",
              "3  TargetLanguage.es  Los grandes modelos de lenguaje son una herramienta poderosa para el procesamiento del lenguaje natural.\n",
              "4  TargetLanguage.he                                                   מודלים לשוניים גדולים הם כלי רב עוצמה לעיבוד שפה טבעית."
            ]
          },
          "execution_count": 5,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "english_text = \"Large Language Models are a powerful tool for natural language processing.\"\n",
        "\n",
        "class TargetLanguage(str, Enum):\n",
        "    de = \"de\"\n",
        "    fr = \"fr\"\n",
        "    it = \"it\"\n",
        "    es = \"es\"\n",
        "    he = \"he\"\n",
        "\n",
        "class Translation(BaseModel):\n",
        "    language: TargetLanguage = Field(\n",
        "        ...,\n",
        "        description=\"The language of the translated text, as 2-letter language code.\"\n",
        "    )\n",
        "    translation: str\n",
        "\n",
        "class Translations(BaseModel):\n",
        "    translations: List[Translation]\n",
        "\n",
        "print_schema(Translations)\n",
        "\n",
        "translations = llm(\n",
        "    Translations,\n",
        "    (\"Translate the user-provided text into the following languages: \" +\n",
        "     json.dumps([language.value for language in TargetLanguage])),\n",
        "    english_text,\n",
        "    temperature=0.7,\n",
        ")\n",
        "\n",
        "pd.DataFrame(translations.dict()[\"translations\"])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Just as we can translate to different languages, we can also use the format to rewrite text in a different style or tone within the same language."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 6,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_c8060_row0_col0, #T_c8060_row0_col1, #T_c8060_row1_col0, #T_c8060_row1_col1, #T_c8060_row2_col0, #T_c8060_row2_col1, #T_c8060_row3_col0, #T_c8060_row3_col1, #T_c8060_row4_col0, #T_c8060_row4_col1, #T_c8060_row5_col0, #T_c8060_row5_col1 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_c8060\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_c8060_level0_col0\" class=\"col_heading level0 col0\" >style</th>\n",
              "      <th id=\"T_c8060_level0_col1\" class=\"col_heading level0 col1\" >text</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_c8060_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_c8060_row0_col0\" class=\"data row0 col0\" >TextStyle.formal</td>\n",
              "      <td id=\"T_c8060_row0_col1\" class=\"data row0 col1\" >Large Language Models are a potent instrument for natural language processing.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_c8060_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_c8060_row1_col0\" class=\"data row1 col0\" >TextStyle.informal</td>\n",
              "      <td id=\"T_c8060_row1_col1\" class=\"data row1 col1\" >Large Language Models are super useful for natural language processing.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_c8060_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_c8060_row2_col0\" class=\"data row2 col0\" >TextStyle.casual</td>\n",
              "      <td id=\"T_c8060_row2_col1\" class=\"data row2 col1\" >Large Language Models are pretty handy for natural language processing.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_c8060_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_c8060_row3_col0\" class=\"data row3 col0\" >TextStyle.academic</td>\n",
              "      <td id=\"T_c8060_row3_col1\" class=\"data row3 col1\" >Large Language Models represent a significant advancement in the field of natural language processing.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_c8060_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_c8060_row4_col0\" class=\"data row4 col0\" >TextStyle.professional</td>\n",
              "      <td id=\"T_c8060_row4_col1\" class=\"data row4 col1\" >Large Language Models serve as an effective solution for various natural language processing tasks.</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_c8060_level0_row5\" class=\"row_heading level0 row5\" >5</th>\n",
              "      <td id=\"T_c8060_row5_col0\" class=\"data row5 col0\" >TextStyle.business</td>\n",
              "      <td id=\"T_c8060_row5_col1\" class=\"data row5 col1\" >Large Language Models offer substantial benefits for enhancing natural language processing capabilities.</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "                    style                                                                                                      text\n",
              "0        TextStyle.formal                            Large Language Models are a potent instrument for natural language processing.\n",
              "1      TextStyle.informal                                   Large Language Models are super useful for natural language processing.\n",
              "2        TextStyle.casual                                   Large Language Models are pretty handy for natural language processing.\n",
              "3      TextStyle.academic    Large Language Models represent a significant advancement in the field of natural language processing.\n",
              "4  TextStyle.professional       Large Language Models serve as an effective solution for various natural language processing tasks.\n",
              "5      TextStyle.business  Large Language Models offer substantial benefits for enhancing natural language processing capabilities."
            ]
          },
          "execution_count": 6,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "input_text = \"Large Language Models are a powerful tool for natural language processing.\"\n",
        "\n",
        "class TextStyle(str, Enum):\n",
        "    formal = \"formal\"\n",
        "    informal = \"informal\"\n",
        "    casual = \"casual\"\n",
        "    academic = \"academic\"\n",
        "    professional = \"professional\"\n",
        "    business = \"business\"\n",
        "\n",
        "class NormalizedText(BaseModel):\n",
        "    style: TextStyle = Field(..., description=(\"The style of the text normalization.\"))\n",
        "    text: str\n",
        "\n",
        "class NormalizedTexts(BaseModel):\n",
        "    normalized_texts: List[NormalizedText]\n",
        "\n",
        "normalizations = llm(\n",
        "    NormalizedTexts,\n",
        "    (\"Normalize the user-provided text into the following styles: \" +\n",
        "     json.dumps([style.value for style in TextStyle])),\n",
        "    input_text,\n",
        "    temperature=0.7,\n",
        ")\n",
        "\n",
        "pd.DataFrame(normalizations.dict()[\"normalized_texts\"])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Unstructured Data\n",
        "\n",
        "One very powerful task we can use an LLM for, is parsing unstructured information into a data structure. Addresses, for example, are often found in documents in incosistent formats, and parsing them into a consistent data strcuture can be very useful for using them in a software system."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 7,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "C9xib4jBkQrx",
        "outputId": "31c103a7-60d5-4d34-818d-479ddcffbd37"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "{\n",
            "  \"first_name\": \"Sherlock\",\n",
            "  \"last_name\": \"Holmes\",\n",
            "  \"street\": \"Baker Street\",\n",
            "  \"house_number\": \"221B\",\n",
            "  \"postal_code\": \"NW1 6XE\",\n",
            "  \"city\": \"London\",\n",
            "  \"state\": \"\",\n",
            "  \"country\": \"United Kingdom\"\n",
            "}\n"
          ]
        }
      ],
      "source": [
        "address_str = (\n",
        "    \"Sherlock Holmes lives in the United Kingdom. \"\n",
        "    \"His residence is in at 221B Baker Street, London, NW1 6XE.\"\n",
        ")\n",
        "\n",
        "class AddressInfo(BaseModel):\n",
        "    first_name: str\n",
        "    last_name: str\n",
        "    street: str\n",
        "    house_number: str\n",
        "    postal_code: str\n",
        "    city: str\n",
        "    state: str\n",
        "    country: str\n",
        "\n",
        "address_info = llm(\n",
        "    AddressInfo,\n",
        "    address_str,\n",
        ")\n",
        "\n",
        "print_result(address_info)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "That was easy! We didn't even have to prompt, just let the LLM know what is the data structure we are expecting. How about a more complex input with multiple addresses? We should be able to get the LLM to process that too."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 8,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_e5ef6_row0_col0, #T_e5ef6_row0_col1, #T_e5ef6_row0_col2, #T_e5ef6_row0_col3, #T_e5ef6_row0_col4, #T_e5ef6_row0_col5, #T_e5ef6_row0_col6, #T_e5ef6_row0_col7, #T_e5ef6_row1_col0, #T_e5ef6_row1_col1, #T_e5ef6_row1_col2, #T_e5ef6_row1_col3, #T_e5ef6_row1_col4, #T_e5ef6_row1_col5, #T_e5ef6_row1_col6, #T_e5ef6_row1_col7, #T_e5ef6_row2_col0, #T_e5ef6_row2_col1, #T_e5ef6_row2_col2, #T_e5ef6_row2_col3, #T_e5ef6_row2_col4, #T_e5ef6_row2_col5, #T_e5ef6_row2_col6, #T_e5ef6_row2_col7, #T_e5ef6_row3_col0, #T_e5ef6_row3_col1, #T_e5ef6_row3_col2, #T_e5ef6_row3_col3, #T_e5ef6_row3_col4, #T_e5ef6_row3_col5, #T_e5ef6_row3_col6, #T_e5ef6_row3_col7, #T_e5ef6_row4_col0, #T_e5ef6_row4_col1, #T_e5ef6_row4_col2, #T_e5ef6_row4_col3, #T_e5ef6_row4_col4, #T_e5ef6_row4_col5, #T_e5ef6_row4_col6, #T_e5ef6_row4_col7 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_e5ef6\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_e5ef6_level0_col0\" class=\"col_heading level0 col0\" >first_name</th>\n",
              "      <th id=\"T_e5ef6_level0_col1\" class=\"col_heading level0 col1\" >last_name</th>\n",
              "      <th id=\"T_e5ef6_level0_col2\" class=\"col_heading level0 col2\" >street</th>\n",
              "      <th id=\"T_e5ef6_level0_col3\" class=\"col_heading level0 col3\" >house_number</th>\n",
              "      <th id=\"T_e5ef6_level0_col4\" class=\"col_heading level0 col4\" >postal_code</th>\n",
              "      <th id=\"T_e5ef6_level0_col5\" class=\"col_heading level0 col5\" >city</th>\n",
              "      <th id=\"T_e5ef6_level0_col6\" class=\"col_heading level0 col6\" >state</th>\n",
              "      <th id=\"T_e5ef6_level0_col7\" class=\"col_heading level0 col7\" >country</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_e5ef6_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_e5ef6_row0_col0\" class=\"data row0 col0\" >Elena</td>\n",
              "      <td id=\"T_e5ef6_row0_col1\" class=\"data row0 col1\" >Martinez</td>\n",
              "      <td id=\"T_e5ef6_row0_col2\" class=\"data row0 col2\" >Elm Street</td>\n",
              "      <td id=\"T_e5ef6_row0_col3\" class=\"data row0 col3\" >142B</td>\n",
              "      <td id=\"T_e5ef6_row0_col4\" class=\"data row0 col4\" >94107</td>\n",
              "      <td id=\"T_e5ef6_row0_col5\" class=\"data row0 col5\" >San Francisco</td>\n",
              "      <td id=\"T_e5ef6_row0_col6\" class=\"data row0 col6\" >CA</td>\n",
              "      <td id=\"T_e5ef6_row0_col7\" class=\"data row0 col7\" >USA</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_e5ef6_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_e5ef6_row1_col0\" class=\"data row1 col0\" >Hans</td>\n",
              "      <td id=\"T_e5ef6_row1_col1\" class=\"data row1 col1\" >Gruber</td>\n",
              "      <td id=\"T_e5ef6_row1_col2\" class=\"data row1 col2\" >Goethestrasse</td>\n",
              "      <td id=\"T_e5ef6_row1_col3\" class=\"data row1 col3\" >22</td>\n",
              "      <td id=\"T_e5ef6_row1_col4\" class=\"data row1 col4\" >8001</td>\n",
              "      <td id=\"T_e5ef6_row1_col5\" class=\"data row1 col5\" >Zürich</td>\n",
              "      <td id=\"T_e5ef6_row1_col6\" class=\"data row1 col6\" ></td>\n",
              "      <td id=\"T_e5ef6_row1_col7\" class=\"data row1 col7\" >Switzerland</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_e5ef6_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_e5ef6_row2_col0\" class=\"data row2 col0\" >Satoshi</td>\n",
              "      <td id=\"T_e5ef6_row2_col1\" class=\"data row2 col1\" >Nakamoto</td>\n",
              "      <td id=\"T_e5ef6_row2_col2\" class=\"data row2 col2\" >Sakura-cho</td>\n",
              "      <td id=\"T_e5ef6_row2_col3\" class=\"data row2 col3\" >2-15-5</td>\n",
              "      <td id=\"T_e5ef6_row2_col4\" class=\"data row2 col4\" >180-0003</td>\n",
              "      <td id=\"T_e5ef6_row2_col5\" class=\"data row2 col5\" >Musashino-shi</td>\n",
              "      <td id=\"T_e5ef6_row2_col6\" class=\"data row2 col6\" >Tokyo-to</td>\n",
              "      <td id=\"T_e5ef6_row2_col7\" class=\"data row2 col7\" >Japan</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_e5ef6_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_e5ef6_row3_col0\" class=\"data row3 col0\" >Catherine</td>\n",
              "      <td id=\"T_e5ef6_row3_col1\" class=\"data row3 col1\" >Dubois</td>\n",
              "      <td id=\"T_e5ef6_row3_col2\" class=\"data row3 col2\" >Rue de la République</td>\n",
              "      <td id=\"T_e5ef6_row3_col3\" class=\"data row3 col3\" >15</td>\n",
              "      <td id=\"T_e5ef6_row3_col4\" class=\"data row3 col4\" >69002</td>\n",
              "      <td id=\"T_e5ef6_row3_col5\" class=\"data row3 col5\" >Lyon</td>\n",
              "      <td id=\"T_e5ef6_row3_col6\" class=\"data row3 col6\" ></td>\n",
              "      <td id=\"T_e5ef6_row3_col7\" class=\"data row3 col7\" >France</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_e5ef6_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_e5ef6_row4_col0\" class=\"data row4 col0\" >David</td>\n",
              "      <td id=\"T_e5ef6_row4_col1\" class=\"data row4 col1\" >Johnson</td>\n",
              "      <td id=\"T_e5ef6_row4_col2\" class=\"data row4 col2\" >Queen Street</td>\n",
              "      <td id=\"T_e5ef6_row4_col3\" class=\"data row4 col3\" >34</td>\n",
              "      <td id=\"T_e5ef6_row4_col4\" class=\"data row4 col4\" >M5H 2Y4</td>\n",
              "      <td id=\"T_e5ef6_row4_col5\" class=\"data row4 col5\" >Toronto</td>\n",
              "      <td id=\"T_e5ef6_row4_col6\" class=\"data row4 col6\" >ON</td>\n",
              "      <td id=\"T_e5ef6_row4_col7\" class=\"data row4 col7\" >Canada</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "  first_name last_name                street house_number postal_code           city     state      country\n",
              "0      Elena  Martinez            Elm Street         142B       94107  San Francisco        CA          USA\n",
              "1       Hans    Gruber         Goethestrasse           22        8001         Zürich            Switzerland\n",
              "2    Satoshi  Nakamoto            Sakura-cho       2-15-5    180-0003  Musashino-shi  Tokyo-to        Japan\n",
              "3  Catherine    Dubois  Rue de la République           15       69002           Lyon                 France\n",
              "4      David   Johnson          Queen Street           34     M5H 2Y4        Toronto        ON       Canada"
            ]
          },
          "execution_count": 8,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "input_text = (\n",
        "  \"During my recent travels, I had the pleasure of visiting several fascinating locations. \"\n",
        "  \"My journey began at the office of Dr. Elena Martinez, 142B Elm Street, San Francisco, \"\n",
        "  \"CA 94107, USA. Her office, nestled in the bustling heart of the city, was a hub of \"\n",
        "  \"innovation and creativity. Next, I made my way to the historic residence of Mr. Hans \"\n",
        "  \"Gruber located at 3. Stock, Goethestrasse 22, 8001 Zürich, Switzerland. The old building, \"\n",
        "  \"with its classic Swiss architecture, stood as a testament to the city’s rich cultural \"\n",
        "  \"heritage. My adventure continued at the tranquil countryside home of Satoshi Nakamoto, \"\n",
        "  \"2-15-5, Sakura-cho, Musashino-shi, Tokyo-to 180-0003, Japan. Their home was surrounded by \"\n",
        "  \"beautiful cherry blossoms, creating a picturesque scene straight out of a postcard. In \"\n",
        "  \"Europe, I visited the charming villa of Mme. Catherine Dubois, 15 Rue de la République, \"\n",
        "  \"69002 Lyon, France. The cobblestone streets and historic buildings of Lyon provided a \"\n",
        "  \"perfect backdrop to her elegant home. Finally, my journey concluded at the modern apartment \"\n",
        "  \"of Mr. David Johnson, Apt 7B, 34 Queen Street, Toronto, ON M5H 2Y4, Canada. The sleek \"\n",
        "  \"design of the apartment building mirrored the contemporary vibe of the city itself.\"\n",
        "  )\n",
        "\n",
        "class Addresses(BaseModel):\n",
        "    addresses: List[AddressInfo]\n",
        "\n",
        "addresses_in_text = llm(\n",
        "    Addresses,\n",
        "    \"Return all the addresses in the user-provided text.\",\n",
        "    input_text,\n",
        ")\n",
        "\n",
        "pd.DataFrame(addresses_in_text.dict()[\"addresses\"])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "The intelligence of LLMs allows them to \"understand\" complex logical and hierarchical structures. Consider the task of converting some information into a knowledge graph. Turning longform text into structure we can work with as part of a system can be of great value, and our LLM can help us achieve that."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 9,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAs8AAAIeCAYAAABEEOI9AAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjkuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/TGe4hAAAACXBIWXMAAA9hAAAPYQGoP6dpAADzcUlEQVR4nOzdd3xTZRfA8d9N0j2gbNmjDKHQllFomQUSljJERQUBQRFwgeD25cX5ylBQkaFMFWTIRmahZW9ahuxR9i6leyS57x+lkUqBliZNac7Xj58PSe499yRtmpPnPvc5iqqqKkIIIYQQQoiH0tg7ASGEEEIIIR4XUjwLIYQQQgiRQ1I8CyGEEEIIkUNSPAshhBBCCJFDUjwLIYQQQgiRQ1I8CyGEEEIIkUNSPAshhBBCCJFDUjwLIYQQQgiRQ1I8CyGEEEIIkUNSPAshRA5FRESgKAoRERFWizly5EgURbFavMKkb9++eHp62jsNIYTIQopnIYRdzJw5E0VR2LNnT5b7b9++TVBQEK6urqxevdpO2TkOs9nMr7/+il6vp0SJEjg5OVGqVCkMBgM///wzqamp9k5RCCEKFJ29ExBCiExxcXEYDAYOHDjA4sWLad++vb1TKtSSk5Pp1q0ba9asISQkhOHDh1O6dGliYmLYuHEjgwcPZufOnUybNs3eqQohRIEhxbMQokCIj4+nXbt2REVFsWjRIjp06GDvlAq9oUOHsmbNGsaPH88777yT5bFhw4Zx4sQJ1q1b98AYRqMRs9mMs7OzLVMVQogCQ6ZtCCHsLiEhgfbt27Nv3z4WLlxIp06dsjyeOff14sWLdO3aFU9PT0qWLMnw4cMxmUxZtk1MTGTYsGFUqFABFxcXatasydixY1FV1bLNM888Q/369bPs9/TTT6MoCsuWLbPct3PnThRFYdWqVQ/Mf+fOnbRv354iRYrg7u5Oy5Yt2bp16z3bbdmyhUaNGuHq6kq1atWYMmVKtvGSk5N5++23KVGiBF5eXnTu3JmLFy+iKAojR47Msu3Fixfp168fpUuXxsXFhTp16jB9+vQH5gtw/vx5pk6dSvv27e8pnDNVr16dwYMHW25HR0ejKApjx45l/PjxVKtWDRcXFw4fPkxaWhojRoygQYMGFClSBA8PD5o3b054eHiWmHfHGDduHJUqVcLNzY2WLVty6NChbPPIyc9dCCHyi4w8CyHsKjExkQ4dOrB7927+/PNPnnrqqWy3M5lMtGvXjsaNGzN27FjCwsL49ttvqVatGoMGDQJAVVU6d+5MeHg4/fv3JyAggDVr1vDee+9x8eJFxo0bB0Dz5s1ZunQpcXFxeHt7o6oqW7duRaPRsHnzZjp37gzA5s2b0Wg0NG3a9L75b9iwgQ4dOtCgQQP++9//otFomDFjBq1bt2bz5s0EBQUBcPDgQQwGAyVLlmTkyJEYjUb++9//Urp06Xti9u3bl/nz5/Pyyy/TpEkTNm7ceM8XCoCrV6/SpEkTFEXhzTffpGTJkqxatYr+/fsTFxfHkCFD7pv3qlWrMJlM9OrV677b3M+MGTNISUlhwIABuLi4UKxYMeLi4pg6dSovvvgir732GvHx8UybNo127dqxa9cuAgICssT49ddfiY+P54033iAlJYXvv/+e1q1bc/DgwSyvSU5+7kIIka9UIYSwgxkzZqiAWqlSJdXJyUldsmTJfbft06ePCqiff/55lvsDAwPVBg0aWG4vWbJEBdQvv/wyy3bPPvusqiiKevLkSVVVVXX37t0qoK5cuVJVVVU9cOCACqjPPfec2rhxY8t+nTt3VgMDAy23w8PDVUANDw9XVVVVzWazWr16dbVdu3aq2Wy2bJeUlKRWqVJF1ev1lvu6du2qurq6qmfPnrXcd/jwYVWr1ap3/yneu3evCqhDhgzJ8hz69u2rAup///tfy339+/dXn3jiCfXGjRtZtn3hhRfUIkWKqElJSdm8mhmGDh2qAmpUVFSW+1NTU9Xr169b/r879pkzZ1RA9fb2Vq9du5ZlP6PRqKampma579atW2rp0qXVfv363RPDzc1NvXDhguX+nTt3qoA6dOhQy305/bkLIUR+kmkbQgi7unr1Kq6urlSoUOGh2w4cODDL7ebNm3P69GnL7ZUrV6LVann77bezbDds2DBUVbVMvwgMDMTT05NNmzYBGSPM5cuXp3fv3uzbt4+kpCRUVWXLli00b978vvlERUVx4sQJXnrpJW7evMmNGze4ceMGiYmJtGnThk2bNmE2mzGZTKxZs4auXbtSsWJFy/5PPvkk7dq1yxIzc4WRu6dLALz11ltZbquqysKFC3n66adRVdVy7Bs3btCuXTtu377Nvn377pt7XFwcwD1Lwa1cuZKSJUta/q9UqdI9+3bv3p2SJUtmuU+r1VrmPZvNZmJiYjAajTRs2DDbPLp27Uq5cuUst4OCgmjcuDErV668Z9uH/dyFECI/SfEshLCrKVOm4OzsTPv27Tl27Nh9t3N1db2nYPPx8eHWrVuW22fPnqVs2bJ4eXll2e7JJ5+0PA4ZhV5wcDCbN28GMorn5s2b06xZM0wmEzt27ODw4cPExMQ8sHg+ceIEAH369MlScJYsWZKpU6eSmprK7du3uX79OsnJyVSvXv2eGDVr1sxy++zZs2g0GqpUqZLlfl9f3yy3r1+/TmxsLD///PM9x37llVcAuHbt2n1zz3yNEhISstzftGlT1q1bx7p16zAYDNnu++/cMs2aNYt69erh6upK8eLFKVmyJH/99Re3b9++Z9vsXosaNWoQHR2d5b6c/NyFECI/yZxnIYRd1a5dm5UrV9KmTRv0ej1bt27NdhRaq9Va9bjNmjXjq6++IiUlhc2bN/PJJ59QtGhR/Pz82Lx5s2Xe7YOKZ7PZDMCYMWPumdObydPT0yZrJWceu1evXvTp0yfbberVq3ff/WvVqgXAoUOH8Pf3t9xfsmRJ2rZtC8Dvv/+e7b5ubm733Pf777/Tt29funbtynvvvUepUqXQarX873//49SpUzl7Utmw9s9dCCHySopnIYTdBQUFsWTJEjp16oRer2fz5s33jDbmRKVKlQgLCyM+Pj7L6PPRo0ctj2dq3rw5aWlp/PHHH1y8eNFSJLdo0cJSPNeoUSPbC/oyVatWDQBvb29LwZmdkiVL4ubmZhmpvtu/R9srVaqE2WzmzJkzWUZnT548eU9MLy8vTCbTA499Px06dECr1TJ79mx69uyZ6/3/7c8//6Rq1aosWrQoS8fE//73v9lun91rcfz4cSpXrpznXIQQwpZk2oYQokBo06YNf/zxBydPnqR9+/aWObm50bFjR0wmExMmTMhy/7hx41AUJcva0Y0bN8bJyYlRo0ZRrFgx6tSpA2QU1Tt27GDjxo0PHHUGaNCgAdWqVWPs2LH3TH+AjKkVkDF62q5dO5YsWcK5c+csjx85coQ1a9Zk2SdzDvTEiROz3P/jjz9mua3VaunevTsLFy7Mdom3zGPfT8WKFenXrx+rVq265/XKpN61vN/DZI4Q373Pzp072b59e7bbL1myhIsXL1pu79q1i507d8r63kKIAk9GnoUQBUa3bt345Zdf6NevH507d2b16tW4urrmeP+nn36a0NBQPvnkE6Kjo/H392ft2rUsXbqUIUOGWEaKAdzd3WnQoAE7duywrPEMGSPPiYmJJCYmPrR41mg0TJ06lQ4dOlCnTh1eeeUVypUrx8WLFwkPD8fb25vly5cD8Nlnn7F69WqaN2/O4MGDMRqN/Pjjj9SpU4cDBw5YYjZo0IDu3bszfvx4bt68aVmq7vjx4wBZRnW/+eYbwsPDady4Ma+99hq1a9cmJiaGffv2ERYWRkxMzAPzHz9+PGfOnOGtt95i7ty5PP3005QqVYobN26wdetWli9ffs+c7Pt56qmnWLRoEd26daNTp06cOXOGyZMnU7t27Wy/WPj6+tKsWTMGDRpEamoq48ePp3jx4rz//vs5Op4QQtiLFM9CiALllVdeISYmhuHDh/Pcc8+xePHiHO+r0WhYtmwZI0aMYN68ecyYMYPKlSszZswYhg0bds/2maPMzZo1s9xXpkwZfH19OXny5EOLZ4BWrVqxfft2vvjiCyZMmEBCQgJlypShcePGvP7665bt6tWrx5o1a3j33XcZMWIE5cuX57PPPuPy5ctZimfIWAO5TJky/PHHHyxevJi2bdsyb948atasmeXLROnSpdm1axeff/45ixYtYuLEiRQvXpw6deowatSoh+bu7u7O6tWr+e233/jtt98YPXo0cXFxFC1aFH9/fyZOnHjf+dT/1rdvX65cucKUKVNYs2YNtWvX5vfff2fBggVERETcs33v3r3RaDSMHz+ea9euERQUxIQJE3jiiSdydDwhhLAXRc3NeTkhhBB2ERUVRWBgIL///rtV5ijbS3R0NFWqVGHMmDEMHz7c3ukIIUSuyZxnIYQoYJKTk++5b/z48Wg0Glq0aGGHjIQQQmSSaRtCCFHAjB49mr179xIaGopOp2PVqlWsWrWKAQMG5KiZjBBCCNuR4lkIIQqYkJAQ1q1bxxdffEFCQgIVK1Zk5MiRfPLJJ/ZOTQghHJ7MeRZCCCGEECKHZM6zEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5JMWzEEIIIYQQOSTFsxBCCCGEEDkkxbMQQgghhBA5pLN3AkJYk2o0Yrp2DTUuDsxmcHFBW6oUGi8ve6cmhBBCiEJAimfx2FPT00k/eJC0vXsxXbmSUTT/i+LujlOdOjg3bIi2VCk7ZCmEEEKIwkBRVVW1dxJCPApVVUnfv5/k1ashNRUUBR7063zncV2tWrh16oTG0zP/khVCCCFEoSDFs3gsqSkpJC1ciPHkydzvrCjg7Iz7M8/gVKOG9ZMTQgghRKElxbN47KgpKSTMnIn52rUHjzQ/jKLg9swzOPv5WS85IYQQQhRqstqGeKyoqkrivHl5L5wzgpG8aBHG8+etk5wQQgghCj0pnsVjJW33bkzR0XkvnO+StHgxanq61eIJIYQQovCS4lk8NswJCaSsW2fdoKqKGhtL6pYt1o0rhBBCiEJJimfx2EiLjASTyfqBVZW0XbtQjUbrxxZCCCFEoSLFswMKCAggICCA2rVro9VqLbd79OhBREQEAQEBeYq/ceNGfH19LbfT0tJwd3dn5MiRlvvmzp1L8+bNcxxTVVXSdu8GVaXuuHEcuHw52+3eWrqUTWfO5DpnNSWF9KNHAYiKimLu3Lm5jvEwI0eOpGTJkgQEBPDkk0/So0cPbt269UixZs6cydE7+QohhBAi/0jx7ICioqKIiopi5cqVeHl5WW7PmzfPKvGbNGnCxYsXuXDhAgA7d+6kXr16REREWLYJDw+ndevWOY6pxsaixsc/dLsfu3ShRZUquc4ZjSZjLjW2K54BevbsSVRUFIcOHSI9PZ0vv/zykeJI8SyEEELYhxTP4h5Go5HBgwfj7+9PnTp12LNnj+WxNWvW0KxZMxo0aEBQUBDh4eH37O/i4kJISIilWI6IiKBv375cunSJlJQUy32hoaGsX7+e4OBgAgMDqVOnDtOmTbPEmTp1KrVr1yYgIIB6jRuz504xDrD8yBH0U6dSb/x4xmzcaLm/04wZrDhyBIBBixczZPlyOs+aRYMffqDX3Lmk3ZmaEZ+ayisLFtDoxx/pMH06Q5Yu5dWvvuLatWuMGDGC8PBwAgICGDhwoOV5169fn3r16tGyZUsOHz5seR5+fn73fb3uR6vV0rZtW44dOwbAmDFjqFOnDnXr1qVnz57cvn0743kuX069evUICAjAz8+PpUuXMnXqVPbs2cPQoUMJCAhg5cqVDz2eEEIIIaxDimdxj6NHj9KnTx/279/PW2+9xSeffALA6dOnGTlyJCtXrmTv3r3MmTOHl156idTU1HtihIaGWgrr8PBwWrVqRePGjdm+fTuXLl3i/PnzBAcHU79+fbZs2UJkZCSbN2/m888/t4xYDxs2jPXr1xMVFcWOX36h1l1ttW+npLDu1VcJf+01fty2jUtxcdk+l4NXrjD3xRfZ9eabXE9MZNmdwnr0xo246nTsevNN5vfsya7z51HT0ihVqhSff/45oaGhREVFMXnyZK5du8ZLL73ErFmzOHDgAAMGDODZZ58lc4n0+71eD5KcnMySJUto0KABq1atYvr06WzdupWDBw/i4eHBhx9+CMCnn37KlClTiIqK4sCBA7Rs2ZJXX32Vhg0bMm7cOKKioujYsWNOf7RCCCGEyCMpnsU9fH19ady4MQDBwcGcOnUKgNWrV3Py5ElatGhBQEAAzz77LBqNhnPnzt0TIzQ0lIiICFJTU4mOjqZWrVq0bNmSiIgIIiIiCA4OxsXFhZs3b/Lcc8/h5+dH69atuXnzJocOHQKgTZs2vPzyy3z//fecuXABT1dXS/zn6tYFoLiHB5V8fDh7n7nDT9WqhbuzM1qNhvrlynEmJgaAjadP0zMwEEVR8HJxoVudOvdd/m7nzp3UrVuXuneO2bNnTy5dusTFixcf+HplZ/bs2QQGBhISEkLt2rX58MMPCQsLo0ePHhQtWhSAQYMGse7OqiJt2rThnXfeYfTo0Rw4cMCyjRBCCCHsQ2fvBETB43pXkarVajHemeqgqip6vZ45c+Y8NEZQUBBXr15l/vz5BAUFAdCyZUv69+/PpUuXLPOdBw4cSMeOHVm4cCGKolC/fn3L1I6FCxeyd+9eIiIi6PrBB3wSHEz3O90AXXT//OpqFQWj2ZxtHjndTlGUjLbdj+B+r1d2evbsyfjx4x8YT7krj++++46///6b8PBw+vTpQ8+ePXn//fcfKU8hhBBC5J2MPIsca9euHWFhYRw4cMBy365du7Ld1snJiaZNm/LFF1/QqlUrAKpXr8758+dZt24doaGhANy6dYtKlSqhKAqbNm1i//79QMa861OnTtGwYUOGDx/OMx07sveuOc951aJKFf6IikJVVRJSU1n8998obm4AeHt7W+YcQ8YFkAcPHrSMiM+dO5dy5cpRrlw5q+TStm1b5s+fT9ydqSdTpkzBYDAAGVNC6tSpw5tvvsmgQYPYsWNHtjkKIYQQIn/IyLPIMV9fX+bMmcPrr79OUlISaWlpBAYG3nckOjQ0lLVr11qKZ8goRFesWGEZjf7mm28YPHgwX3zxBQEBAZbpDyaTiX79+hETE4NOp6NEsWL8EBxstefyfsuWvLVsGUETJlDc3R2/MmXwKVkSyJgqMXbsWOrVq0dISAiTJ09m9uzZ9O7dG6PRiI+PDwsWLMgyQpwXHTp04NChQwQHB6PRaKhXrx4TJ04E4OOPP+bYsWM4Ozvj7u7OpEmTABgwYADDhg1j3LhxfP311zLvWQghhMgniqpasc+xEDYUP3ky5mvXrNKaO91kwmQ24+rkRGJaGt1/+4233nuPl956ywqZCiGEEKKwkuJZPDbSIiNJXrbMKrGuJyTw7OzZmMxmUo1GOgUE8O3atWg0MpNJCCGEEPcnxbN4bKjp6cT/8ANqYqJVRp/v5ta5M86BgVaNKYQQQojCR4bZxGNDcXLCrUsX6xbOGg3aypVxymNLciGEEEI4BimexWPFydcX5zsXG+aZoqC4uODepYvVLv4TQgghROEmxbN47Li2a4dTvXp5C6LRoLi44NG7NxppPCKEEEKIHJI5z+KxpKoqqVu2kBoRkXHHfZqf3I+2bFncundHW6yY9ZMTQgghRKElxbN4rJmuXSN55UpMZ8+CRnP/IlpRQFVR3Nxwad4c58aNUWRlDSGEEELkkhTPolAwXb9OWmQkxrNnMV+9CiaT5THFy4vUYsXwDAjA1c8PRSe9gYQQQgjxaKR4FoWOajajJieD2Yzi7AzOzri6ulKhQgWOHDmCk5OTvVMUQgghxGNKzluLQkfRaNB4eKDx8kJxceHixYukpaVx6tQp+vXrh3xfFEIIIcSjkuJZFHp79uyx/Pv333/n448/tmM2QgghhHicSfEsCr29e/eiu2ue8zfffMOECRPsmJEQQgghHldSPItCb9euXZjuuoAQ4O2332bVqlV2ykgIIYQQjyspnkWhpqoqu3fvvmees6urKzdv3rRTVkIIIYR4XEnxLAq1pKQkYmNjURQFX19fAGbPns3t27fp1auXnbMTQgghxONGFrwVhZqHhwcHDhygXLlyODs74+Pjw82bN2W5OiGEEEI8ElnnWTiU1q1b4+npybJly+ydihBCCCEeQzJtQzgUvV5PeHg4aWlp9k5FCCGEEI8hKZ6FQzEYDCQkJLBjxw57pyKEEEKIx5AUz8KhBAYGUrx4cdatW2fvVIQQQgjxGJLiWTgUjUZD27ZtWbt2rb1TEUIIIcRjSIpn4XAMBgO7d+8mJibG3qkIIYQQ4jEjxbNwOHq9HlVV2bBhg71TEUIIIcRjRopn4XAqVKhArVq1ZOqGEEIIIXJNimfhkAwGA2vXrr2nbbcQQgghxINI8SwcksFg4OzZs5w4ccLeqQghhBDiMSLFs3BILVu2xMnJSZasE0IIIUSuSPEsHJKnpychISEy71kIIYQQuSLFs3BYBoOB8PBw0tPT7Z2KEEIIIR4TUjwLh2UwGIiPj2fnzp32TkUIIYQQjwkpnoXDCgwMpFixYjJ1QwghhBA5JsWzcFharVZadQshhBAiV6R4Fg4ts1X3rVu37J2KEEIIIR4DUjwLh6bX6zGbzdKqWwghhBA5IsWzcGgVK1akZs2aMnVDCCGEEDkixbNweNKqWwghhBA5JcWzcHh6vZ7o6GhOnTpl71SEEEIIUcBJ8SwcXqtWrdDpdDJ1QwghhBAPJcWzcHheXl7SqlsIIYQQOSLFsxBIq24hhBBC5IwUz0KQMe85Li6OXbt22TsVIYQQQhRgOnsnIMTdVJMJ8/XrmBMSwGxGcXVFW7o0iouLTY/boEEDfHx8WLt2LU2bNrXpsR5EVVXU2FhMMTFgMqHodGhKlEDx8kJRFLvlJYQQ+SVVTeWG6QZpahoKCh6KB8W0xdAqWnunJgQAiirrcwk7U9PTST90iLR9+zBdugRm8z3bKEWL4ly3Ls4NGqApUsQmeTz//PNcuHCBbdu22ST+/aiqiunMGVL37MF4+jSkpt6zjeLmhq5GDZwbNkRbrpwU0kKIQiXeHM+h1EMcTTtKnDnunsc1aCilLYWfix81nWuiU2TsT9iPFM/CblRVJT0ykuS1azMKRkWBB/063nncqX593AwGq49G//LLLwwcOJCbN29StGhRq8a+H+OFCyQvXYr5xg3QaLL94mBx53Ft+fK4demCtkSJfMlRCCFsJVVNZUvSFg6lHUJBQeXhJYkzzrRwb0Ft59oykCDsQopnYRfm5GSS/vwT0+nTud9ZUVA8PHB//nl0FSpYLafo6GiqVKnCwoULeeaZZ6wWNzuqqpIaHk7q5s0P/9Lwb5qMSxVc27XDJSjIRhkKIYRtXTJeYmXCSpLUpBwVzf9WUVeRDh4dcNW42iA7Ie5PLhgU+c6clETi9OmYzpx5tACqipqYSOKsWRnTHKykcuXK1KhRg3Xr1lktZnZUVSV52bKMwjnjjtwFMJvBbCZl1SpSIiKsnp8QQtjaufRzLIpf9MiFM8B543kWxC8g2Zxs5eyEeDApnkW+Us1mkmbPxnzzZu6LxiyBVDCbSfzjD0zXr1stP71eb/P1nlMjIkiPirJOrI0bSdu3zyqxhBAiP8SYYliWsAwTpkcunAFUVG6Zb7E0YSlm9QFT3oSwMimeRb5K3bYt46JAa8wWulNAJy9ejPqgucK5YDAYOH36tM1adRsvXiR10yarxkxetQpzbKxVYwohhC2YVTNrEtdgxjp/s1VUrpqusi9VBhFE/pHiWeQbc2wsqeHhVg5qxnT5Mml79lglnC1bdauqSvLSpRlznK3JZCL5r7+sG1MIIWzgYOpBrpmu5WnEOTvbk7cTZ7p3lQ4hbEGKZ5Fv0vbssc6IczZSt2/HGte+ent706RJE5vMezZFR2O+ft36r4GqYjx5MmNtaCGEKKBUVbXZCLGKysG0gzaJLcS/SfFcwAUEBBAQEEDt2rXRarWW2z169CAiIoKAgIA8xd+4cSO+vr6W22lpabi7uzNy5EjLfXPnzqV58+YPjNO3b1/Gjx8PwMiRIxkyZEiWx1WTidS9e9ly5gz6qVNpNmkSjSdMoN20aVxLSABgdmQkxx9x/rIaG3vfCxCjo6OZPHnyfff99+toMBhYv349RqPxkXK526JFi2jQoEHGz7BZM56eNQuz2UzdceM4cPlytvs86LH7UhTS9u7Nc75CCGEr543nGVZ3GF8FfcXoFqMZ3WI0c9+ee892c9+ey4nNJ3IVW0XlYOpBTKrJWukKcV+yyngBF3XnwrLo6GgCAgIstyGj6MurJk2acPHiRS5cuED58uXZuXMn9erVyxI7PDyc1q1b5+k45mvXMCYm0mvuXJb07k1A2bIAnLhxA3cnJwDmREVRxNWVGiVLZhvDZDaj1dzn+55GgzE6Gl3Vqvc8lFk8Dxw4MEe5GgwGRowYwa5duwgJCcnRPv9mNBq5fv06AwYMYO/evVSsWJG40aOJOn3aNuuSqirGU6dAr7d+bCGEsIILxgsoKPSZ1ofydctnu43ZZOaFH154pPipaioxphhK6rL/DBHCWmTk+TFnNBoZPHgw/v7+1KlThz13zf1ds2YNzZo1o0GDBgQFBRGezXxjFxcXQkJCLMVyREQEffv25dKlS6SkpFjuCw0NZf369QQHBxMYGEidOnWYNm1ajvM0Xb5MQloa8amplPb0tNxfvUQJPF1c+HXvXqIuXeLjNWtoNmkSa48fZ3ZkJE/NnMnL8+YRMnEiey9eZN/Fizw9cyatpkyh+eTJLPn774xAZjOrV63K9vkOHDiQY8eOERAQQOfOnR+a69GjR9FoNHTr1o1OnTpx8eJFAGbOnEnXrl0t261YsYJWrVpZXqM6derQv39/AgICWLx4MQ0aNCA5OZnnnnuOqpUrM2bNGgLKlkVRFOJTU3l+9mye+OorSn3xBe8sW2aJazSbGbp8OSETJ9Js0iRWHDlieex+z/9GYiJdxo6lbt261KtXj1deeSXHPxshhMgPV41X75nrvHPOTiZ0nsD03tMZ1XQUZ/ee5cenf+TAXwcASIlPYe47c/mu7XeMajaKeUPmYUzLOCu4duxavm78tWUUO+Z8DFdNV/P9eQnHIyPPj7mjR48ybdo0Jk6cyOTJk/nkk09Ys2YNp0+fZuTIkaxZswZvb29OnjxJ8+bNiY6OxuVfnflCQ0MJDw+nV69ehIeHM3HiRLZu3cr27dupWbMm58+fJzg4mKSkJLZs2YJWqyUmJobAwEDatWtH+fLZjyDczRwTQ1EPD14LCqLRhAkEV6xIowoVeKZOHXxLlKB3gwbMO3CAQU2a8NSTTwIZ0zj2XrjApoEDqV6iBLHJyTw9axYLevakjJcXNxMTaTllCkEVKpBmNPL10qWEHThwz/OdPHkyQ4YMyTJqfz+HDh3igw8+oH379sTGxhISEsKrr77KqlWrHrrvkSNHmDhxouVLxXvvvccTTzzByZMnaezvz9jNm2nr60tguXK4Oznx1JNPMrpjR8JOnOD5OXP4oFUrynp7czMpiZ4BAXzapg2nbt6k7dSp1HviCbxdXHhn+fJsn//iv/+mUtGihK1bh8bHhxiZ/yyEKGBumW8BMKv/LJxcM844Br0QxNm9ZxkeMZzS1Uvfs8/S/yylanBVXvj+BVRVZd4789g0eRNNejdhw4QNfH7kc5zdnElLSkOr0XLbfDtfn5NwTFI8P+Z8fX1p3LgxAMHBwYwdOxaA1atXc/LkSVq0aGHZVqPRcO7cOapXr54lRmhoKDNmzCA1NZXo6Ghq1apFy5YtiYiI4PLlywQHB+Pi4sL58+fp378/x48fR6fTcfPmTQ4dOpSj4lk1ZcxD+6ZDBwYHB7P5zBk2njlDiylTWNirF8GVKmW7X1CFClS/04Z61/nznL11i2d//z3LNidu3ODEjRucvnEj2+ebG+Hh4bRv356QkBAGDx7MnDlz+PzzzzGZHj6PrmrVqrRs2TLLfb///jtFixZlw4IFbNu+nU4zZ7Jl4EDSTSa2nztH8E8/odVoMKsqx69fx8vFhTSTiU61agFQrXhxmlSsyPazZ/Fxc7vv829UvjyTtm9n2Cef0KpdO9q3b5+r5y2EELaWOR/57mkbO+fspEpQlWwLZ4CDKw8SvTuaiIkRAKSnpKNoFVy9XClZrSS/v/47NVvXpI6+DsXKFZM5zyJfSPH8mHN1/actqVartVzkpqoqer2eOXPmPDRGUFAQV69eZf78+QTdaffcsmVL+vfvz6VLlyzznQcOHEjHjh1ZuHAhiqJQv359y9SOh1F0//yqVSxalJ6BgfQMDGTo8uUs/vvv+xbPHs7Oln+rQK2SJVn76qv3bHf8xg1Ca9ZkQTYXzWVOu8gNvV6PyWRiy5Ytlvt0Ol2WIvrfz93zrukomVxdXalVqxa+vXvzy6RJaBSFZUeOcCMpiS/btaOHvz9xKSlU/OYbku9zgWLmDOkHPX+AzQMHsqNWLRYtWsR//vMfIiMj0Wq1uXviQghhIzol+5LD2cM52/sh47PslVmvUMq31D2PDV07lDO7znByy0nGGcbRd2pfGoQ2sFq+QtyPzHkupNq1a0dYWBgHDhyw3Ldr165st3VycqJp06Z88cUXljm81atX5/z586xbt47Q0FAAbt26RaVKlVAUhU2bNrF///4c56MpWZKE5GTWnThhWVIuOT2dYzduUKVYMQC8XFyIS029b4zGFSpwNjaWiLsamBy4fJk0o5E2vr5EnDyZ7fP19vbm9u2cncoLDQ1l9erVuLi44Ovry48//kibNm3QarX4+vpy4MABkpOTMRqND/1iYjKZLFNFNCVKYFJVriQkUK5IEVSwzP3++a6fi5eLC85aLSuPHQPg9M2b7Dh3jpBKlR74/KNv3cLD3Z0effrw448/cvz4cRLurGIihBAFQXFNcRRyd8F03Y51Wf/9ekzGjIGLpNgkrp++Tkp8CvHX4qkWXI1277WjapOqnDtwjmKaYrZIXYgsZOS5kPL19WXOnDm8/vrrJCUlkZaWRmBg4H0LvtDQUNauXWspniFjJY4VK1ZYRqO/+eYbBg8ezBdffEFAQIBlukhOaJ94AhWYvmcPH65ahauTE0aTiTa+vrzWqBEAfRs04NO1a5m4fTsj2rS5J0ZRNzfmvfQS/1m7lk/WrsVoMlG+SBFmv/ACVYsXZ9Z//5vt861Xrx516tTBz8+PqlWrsuyui/Mg46LLzBF8Pz8/xowZQ/v27YmJieHSpUscP37c8np07NgRPz8/nnjiCZo2bcrOnTvv+5xVVWXKlCl8/fXXuLu7c+LGDZ6rW5fn6tZl+F9/8frixZTx8qK7n1+W/Yq7u7Ph5ElWHjuGAvzQuTMVihYFuO/z3xIdzcQ//8Rp8WKMRiNjxoyhSJEiOf75CCGErZXWZT8140G6fdWN5Z8vZ0yLMSgaBa1Oy9Mjn8bJxYkZfWeQlpQGCpSsWpKgF4Mopbt3hFoIa1NUa3SWEOIhVFUlfvx41DjbdYDyeO01dHeWwMuNMWPGcPDgQX799dcs9y9dupSuXbty6tQpqmazBF5uJS1fTnpUFFiplXgWioJLq1a43jXnWwghCpIrxivMi59ns/ieiif9ivSzzXKgQtxFpm2IfKEoCi6NGlm/NXVGcDRlyjxS4dyyZUv+/PNPPv3003sea9WqFVqt1mrdBl0aNbJN4QygKDjXr2+b2EIIYQVldGUoqS2Z66kbOeXv6i+Fs8gXUjyLfONUvz7caYhiVaqKa7Nmj7Trxo0b2blzJzVq1LjnsSJFitCkSRPWrl2b1wwB0JYpg7ZyZet/gVAUnOrVQ5PNBYtCCFGQNHJtdM9az9bgjDN1nOtYPa4Q2ZHiWeQbjbs7bh07WjeooqCrWRNd7drWjXuHNVt1A7h37gxWXgFDcXPDzWCwakwhhLAFXydfquiqWH30OdQ9FDeNm1VjCnE/UjyLfOVUr15GoWuN0VdFQXF3x+2pp2x2qk6v13P79u0snRvzQuPjg1uHDlaJBYCi4NatG4qbfGgIIQo+RVFo49EGV8XVKgW0goKvky81nWtaITshckaKZ5GvFEXBvVs3dFWr5q2AVhQUNzc8+vSx6XSFRo0aUaRIEatN3QBwrl8f17Zt8xZEUTIK565dcfL1tU5iQgiRDzw0HnT36m6VArqiriLtPNrJXGeRr2S1DWEXqslESlgYaTt2ZBSCufw11JYvT9+5c/EsV46xY8dS4k4XQlvo3r07V69ezdIwxRrSDh4kecUKMBpzdyFh5oi7FM5CiMfYbdNtVieu5orpSq72U1BQUQl0CaSpW1O0ijSDEvlLimdhV8Zz50j+6y/M166BRnP/IvJOga24uuLSsiXOQUEUKVqU+Ph4nJyc6NevH++++262F/7l1ZQpU3jjjTeIiYnB29vbqrHNcXEkr16N8ciRh3+JuDPa7BQQgFvbtjJVQwjx2DOrZvan7mdnyk5S1VRLYZydzMeKa4rT2qM1ZXW5X2FJCGuQ4lnYnaqqmC5eJC0yEtPZs5hjYrIUkYqHB9py5XDy88PpySctrb779OljWZtZq9ViNpvp1KkTH330ESEhIVbL7/Tp01SrVo0lS5bQpUsXq8W9m/n2bdL27sV4+jSmq1czRqMzOTlBqVJQuTJewcFoPDxskoMQQtiLUTVyMu0kx9OOc8V0hWQ1OcvjPhoffNJ8aOjTkDLaMjJNQ9iVFM+iwFHT01ETEzMKaBcXNO7u2W43ffp0+vfvf8/9bm5uxMbG4uzsbLWcfH19adeuHT/99JPVYt6PajajJiSgGo0oTk4onp7UqFGDy5cvEx0dbdMpKkIIURAkm5NJU9NQUHDTuDFr2ixee+01pkyZwoABA+ydnnBwcsGgKHAUJyc0RYui8fG5b+EMEBwcnHU/RcHV1ZXffvvNqoUzZCxZZ61mKQ+jaDRovL3RFiuGxssLRVG4fPkyiYmJtG/fnqSkpHzJQwgh7MVN40YRbRG8td44KU4cP34cgMGDB7N582Y7ZyccnRTP4rFVs2ZNvLy8ANBoNKiqykcffUT37t2tfiy9Xs+JEyc4c+aM1WM/zKVLl0hMTAQgMjKS559/3mrrTgshxOPg0KFDAJhMJjp16sTff/9t54yEI5PiWTy2NBoNTZs2BSAkJISePXvy1VdfsX37dqsfKzQ01KqtunNj7969ln+bzWZWrlzJoEGDkBlXQghHsWvXLsu/k5KSaNu2LRcuXLBjRsKRSfEsHmvvvfcen332GRs2bGD69Ok0atSIZ555hosXL1r1OEWLFqVx48ZWXe85p/bu3YvuzkWSkHGB5dSpUxk9enS+5yKEEPnt0qVL3Lx503LbZDJx9epV9Ho9ycnJD9hTCNuQ4lk81lq3bs2IESNwcnLC2dmZhQsX4uTkRNeuXa3+R1Wv17N+/XpMJpNV4z7Mrl277pmmoSgKly5dytc8hBDCHv7d4VVRFFRVJSkpiYSEBDtlJRyZFM+iUCldujRLlizh77//5rXXXrPq1AaDwUBsbKzVWnXn1KlTpwAoWzZjTdN33nmHmzdv8v333+drHkIIYQ+Z15q4urpSqlQpqlatyuHDh4mOjqZkyZJ2zk44IimeRaFTv359ZsyYwezZsxk7dqzV4gYFBeHt7Z3vUzdWrFjByZMnuXjxIv7+/ty6dQsfH598zUEIIeylX79+7Nixg9u3bzNixAjOnTtHhQoVZK1nYTdSPItCqUePHnz00Ud88MEHrFq1yioxdTodrVu3zveLBqtXr061atWAf5bMk4sFhRCOwsvLi8aNG+Ps7Ixeryc9PZ2NGzfaOy3hwKR4FoXWl19+SadOnXjxxRc5duyYVWIaDAa2b99OXFycVeLlll6v5/Lly7JMkxDCIVWvXp1KlSrZ5eJtITJJ8SwKLY1Gw+zZsylbtiydO3cmNjY2zzENBgNGo5GIiIg8x3oUzZo1w9XVVT44hBAOSVEUDAaD/A0UdiXFsyjUvL29Wbp0KdeuXePFF1/M80oZ1apVo2rVqnZZ7xkyWo+3aNHCbscXQgh70+v1HD16lPPnz9s7FeGgpHgWhV716tWZN28ea9eu5cMPP8xzPL1eb9dRD71ez8aNG0lJSbFbDkIIYS9t2rRBURQZRBB2I8WzcAgGg4GxY8cyduxYfvvttzzHOn78OGfPnrVSdrk/fnJyMlu3brXL8YUQwp6KFStGo0aNZOqGsBspnoXDGDJkCH369OG1117L0uo1t1q3bo1Go7HbqEfdunUpXbq0jLoIIRyWXq8nLCwMs9ls71SEA5LiWTgMRVGYPHkyAQEBdOvWjcuXLz9SnKJFixIUFGS3UQ9FUew+dUQIIezJYDBw8+ZNIiMj7Z2KcEBSPAuH4urqyuLFiwHo1q3bI88bNhgMhIWF5Xur7ruPHxkZybVr1+xyfCGEsKcmTZrg6ekpgwjCLqR4Fg7niSeeYMmSJURFRTFw4MBHajhiMBi4desW+/bts0GGD9e2bVsA1q9fb5fjCyGEPTk7O9OqVSuZvibsQopn4ZAaNWrE1KlTmTVrFuPHj8/1/kFBQXh5edlt1OOJJ56gbt26MuoihHBYBoOBLVu2kJiYaO9UhIOR4lk4rF69evHee+8xfPjwXBehTk5OtG7d2q7Fa2ajAGnVLYRwRAaDQVp1C7uQ4lk4tP/973+0a9eOHj16cOLEiVztazAY2LZtG/Hx8TbK7uHHv3TpEkeOHLHL8YUQwp5q1KhBxYoVZeqGyHdSPAuHptVqmTNnDqVLl6ZLly7ExcXleN/MVt32GvVo3rw5Li4uMnVDCOGQZOUhYS9SPAuHV7RoUZYuXcqlS5d46aWXcryCRrVq1ahcubLd/nC7ubnRvHlz+eAQQjgsg8HA4cOHuXDhgr1TEQ5EimchgJo1a/LHH3+wcuVKPv300xztoyiKZd6xvRgMBiIiIkhNTbVbDkIIYS/SqlvYgxTPQtzRoUMHRo0axTfffMMff/yRo30MBgPHjh3j3LlzNs7u/sdPTk5m27Ztdjm+EELYU/HixWnQoIEUzyJfSfEsxF2GDx9Or1696NevH3v37n3o9gWhVXepUqVk6oYQwmEZDAbWrVsnrbpFvpHiWYi7KIrCzz//jJ+fH127duXKlSsP3N7Hx4dGjRrZrXjVaDRywYwQwqEZDAZu3LhBVFSUvVMRDkKKZyH+xc3NjSVLlmA0GunevftD5xMXlFbd169ft8vxhRDCnoKDg/Hw8JCpGyLfSPEsRDbKlSvH4sWL2bNnD2+88cYDG5Ho9XpiYmKIjIzMxwz/0bZtW1RVlVbdQgiHlNmqW87AifwixbMQ99GkSROmTJnCtGnTmDBhwgO38/T0tNsf7rJly+Ln5ycfHEIIh5XZqjspKcneqQgHIMWzEA/Qt29fhgwZwtChQ+87spvZqtuepwwzL5iRVt1CCEdkMBhIS0tj06ZN9k5FOAApnoV4iDFjxtC6dWuef/55Tp8+ne02BoOBrVu3kpCQkM/ZZdDr9Vy4cIGjR4/a5fhCCGFPNWvWpHz58nIGTuQLKZ6FeAidTsfcuXPx8fGhc+fOxMfH37ONXq8nPT3dbq26W7RogbOzs3xwCCEcUkFoWiUchxTPQuRAsWLFWLZsGefOnePll1++Zz3R6tWrU6lSJdauXYv59m2MZ89iPH0a4/nzqMnJNs/P3d2d5s2by9XmArNq5qbpJufTz3M+/TzXjNcwqkZ7p5VvVFUlzhzHhfQLnEs/x2XjZVLN0oEzP6WpaVwxXuFc+jkupF8g1hSbL1PKDAYDf//9NxcvXrT5sYRjU1SZJClEji1fvpwuXbrw6aef8vnnnwMZH9amM2fYPmUKlXU6vJ2d79lPKVIEp5o1cW7UCG2JEjbJbdSoUXz++efExMTg4uJik2OIgildTedY2jEOpx7mmukaJrIum6igUFxbnCedn6S2c21cNa52ytQ2VFXlnPEcB1MPcsF4gVT13mLZS+NFNadq1HOph4/Wxw5ZFm7x5ngOpR7ieNpxYs2x9zzuhBNldWXxc/GjqlNVNIr1x+5u3LhBqVKlmDFjBn369LF6fCEySfEsRC7973//4+OPP2b+/Pl0a9KE5GXLMN+4gaooKA96OykKqCq66tVxe+opNN7eVs0rMjKS+vXrEx4eTqtWrawaWxRMqqpyIPUAW5O3kk56jvbRoqWBawOCXIPQKlobZ2h7F9MvEpYURqw5FgUFlfu/BzMfr+pUldburfHQeORjpoVTqjmVTcmbOJx2OMevv4fiQWv31lR1rmr1fBo2bEjNmjWZPXu21WMLkUmKZyFySVVVer70En63bzO4cWOUO0Vxjmk0oNPh9vTTOPv5WS0vs9lMmTJleO211/jqq6+sFlcUTInmRFYlruKi8dFOURfTFKOTZyeKaYtZObP8YVbNbEneQmRq5EOLtn9TUHDCCb2HHl9nXxtmWbhdSL/AqsRVJKvJuXr9M9V0qkkbjzY4KU5Wy+njjz9m6tSpXLlyBY1GZqYK25DiWYhcUs1m4hcswHzkSEbhnAeunTrh0rChlTKDl156iRMnTrB7926rxRQFT4I5gQXxC4g3xz9S0QIZBaSz4kx3z+6U1JW0coa2ZVbNrEpcxcn0k3mO1da9LXVc6lghK8dyJv0MKxJWoN7571EoKJTRlqGbVzerFdDh4eG0bt2ayMhIAgICrBJTiH+Tr2VC5FLK2rWoR4/muXAGSPnrL9KPHbNCVhkMBgN79+7l5s2bVospChajamRR/KI8Fc4AKippahqLEhaRaE60Yoa2tzFpo1UKZ4CwpDCi06OtEstRXDNeY0XCCsyY8/w7eMV0hVUJq6x2QWFISAju7u6y6oawKSmehcgFY3Q0aTt3WjVm8tKlmK3UFUuv10ur7kJue/J2bplv5aloyaSikqqmsiFpw2PTYOdc+jkOpB2wasy1iWtJMadYNWZhZVJNrElcY5XfP8j4HTxjPMORtCNWiefi4iKtuoXNSfEsRA6pZjNJS5dmXPhnzbgpKaSEhVklVrly5ahdu7Z8cBRS143X2Ze6z6oxVVROp5/mdHr2DYAKEpNqYl3iOhSs+x5MUVPYlrzNqjELq30p+4gxx1iteM4UkRRhtS8w0qpb2JoUz0LkkPH4cdTY2NxdHJgTqkr6/v1WG33ObBTwuIwkipyLSo2yeuEIGXNPrV2U28Lp9NMkqAlWL9xUVA6nHZbR54cwqSYiUyNtEjuddA6nHbZKLL1eT2pqKps3b7ZKPCH+TYpnUehVrlyZWrVqYTT+0yiiYcOGRERE5CpO6u7dVh91tlBV0iOt86FkMBg4f/48x48ft0o8AZ999hmvvvqq5faWLVtQFCXL79DAgQP5z3/+A2R0O4uNjc02VseOHTl2Z577zJkzc9xSPVVN5VjaMUvhaDKaWD1qNV83/ppvQr5hdIvRzBsyj6TbSZzYcoLRLUbn+PmpqFwyXiLGFJPjfTKNHDmSIUOGPHS7iIgIVq9eTeXKlYmKisr1cQD2p+6/58tD/PV45rwxhy8Cv2BMyzGMbTWWdd/lvlmQCZPVpg48ripXrkzNmjUJCAjgySef5KWXXiIx8Z/58GfSz5Cs2q7pU1RqlFW+9D/55JOUK1dOzsAJm5HiWTiE1NRUpk2b9sj7q2YzprNnrT/qbDmASvpp65w2DwkJkVbdVhYaGpqlUA4PD6dx48b33Ne6deuHxlq5ciU1a9YEclc8XzFeydL8ZO7bczkfdZ4ha4bw4bYPeW/je9RsVZOkW49+BuN8+vlH3vdhMovnR2VUjVwyXsoy6pyWnMaPT/2ITwUfPtnzCe9tfI+3V76Ns/u9jYrMZvM9nUH/zZbP/3Exb948oqKi+Pvvv7l9+zYzZ860PHbeeB7NQ8oGk9H0wMcfJN4cT4KaAJBlsCO3Mlt1S8dVYStSPAuHMHLkSL744ot75sDFx8fz2muvERQURL169RgwYABpaWkcP36cGjVqABnrOpcpXZrP16wBYGt0NE/f+UD5de9eGk+YQLNJkwiZOJE9Fy4AUHfcOP6zdi0tp0wh8Pvv+WHrVssxP12zhtCff6bZpEl0mD6dEzduAGC6dAlFUfj0008JDAykRo0aWRb63717N61bt6Zhw4YEBgayYMECAKKjoylatCgffPAB9evXZ8aMGTRt2lSKZytq0qQJly5d4sKdn29ERAQjRoywFM+XL1/m3LlzBAcHW/aZOHEiQUFBVKlShRkzZljuzxx5nTp1Knv27GHo0KEEBASwcuVKAMaOHUtQUBD169enffv2nD17FoCrxquWUdfrp68TtTSKFye8iHtRdyCjYAjoGkCJyhkdLM1GMwuGL2B089F8E/wN5yLPWXLYPW83o5qNYlSzUfzc42diL2U0GLlmusaoUaOoW7cu/v7+NGnSxPKeGTNmDHXq1KFu3br07NmT27dv3/M6HTx4kGbNmlG/fn1q167Nl19+CUBUVBSTJ09m9uzZXLp0iSlTpgAwe/ZsSpQogZubG+7u7rz88stA9u/Ly8mX75muse/Pfbh6udLhww5otBkfZ87uzrQc2BKAVd+sYnrv6UzqPolRIaOIuxLHkfVH+L7D94wNHct3bb/jxOYTAMRdjeO9ju/RoEED6tSpw5tvvmkptmfOnEnbtm158cUXqV27NiEhIRw+fJhu3brx5JNPYjAYSEhIyNHv0uMiLS2NpKQkfHx8mDlzJl27duWK8QpmzPy95m9+fPpHAE5sOcE3wd/wx1t/MLrFaA6uOMiZnWcY3WI0o5qOYs6bcxjdfDQntvzzOs98ZSbftf2OUU1H8ddXf1mO+Zn/Zwx/fzhBQUH06dOHEydO0LRpU/z9/albty6ffvppjvPX6/UcPHiQy5cvW/eFEQIpnoWD8Pf3JzQ0lHHjxmW5f9iwYTRv3pxdu3axf/9+zGYz33//PTVq1CA1NZVz585x4MABqpQrx8YzZwAIP32aVlUzOmN9unYtS/v0YcugQWx8/XVqlfxnvdxrCQlEDBhA2Kuv8vPOnew8l1G8DGnWjPABA9gyaBCvNmrEh6tWZeyQkjHfUlEUIiMjWb16NW+99RbR0dHExsYyYMAAZs+ezZ49e1i3bh3Dhg3j4sWMBhm3b9+mTp067Nu3jyFDhmAwGAgPDyctLc2mr6ujcHZ2JiQkhPDwcFJTUzlz5gwdO3bkwoULpKSkEB4eTnBwMK6u/7S9dnFxYdeuXaxatYq33377npG0V199lYYNGzJu3DiioqLo2LEjc+bM4dixY2zfvp19+/bRs2dPBg8eDJCl5fGF/RcoWbUknsU975vztRPXaPRCI97f/D7NBzTnry8zipTLhy+z7L/LeH3+63yw5QMqB1Vm7jtzUVFZ/NtiFi5cyJYtW9i/fz+rVq3CxcWFVatWMX36dLZu3crBgwfx8PDgww8/vOeYlStXZv369ezbt4+9e/eycOFCduzYQUBAAAMHDqRnz56ULVuW119/ndOnTzN48GDeeOMNkpOTOXDgAGvXriU1NTXb9+WP3/94z/HO7z9P5YaVH/izi94TTa9Jvfhox0cY04ysHrWa1+e9zvDw4bz888v8NuA3jKlG3Iq40X9Of3bs2cGBAweIjo5m/vz5lji7d+9m1KhRHD58mGrVqvH0008zefJkjhw5grOzM7NmzXpgHo+LHj16EBAQQJkyZdBoNDz//POWx7Jruw1w9fjVjN+1Te/j19GPWf1n0e2rbnyw9QMa9WjEpb8vWbadPXg2zV5txrth7zJ843DOR54nakkUkDH3/urNq+zcuZPZs2czYcIEnnrqKfbv38/Bgwd59913c/w82rZtCyCjz8ImdPZOQIj88sUXXxAUFMTAgQMt9y1ZsoTt27fz3XffAZCcnIxWm9GyuE2bNoSFhXHr1i1e6tSJX6ZNIzY5mY2nT/O/9u0BaFGlCq8vWkT7mjXR+/riW6KEJfbL9eujKArFPTx46skniTh9msYVKxJ+6hQ/79pFQmoqZlXlVnLWOYSZc2urVq1KixYt2LRpEyVKlOD06dN06NAhy7bHjh2jatWqODk50atXL8v9BoOBjz76iB07dtCiRQsrvoqOK3PqRqVKlQgKCgIyRqS3b99OREQEoaGhWbbv2bMnALVq1UKn03HlyhXKly//wGMsWbKE3bt306BBAwBMpn9OgZvU3J0OL1G1hKWwrNyoMuETwoGMkcJabWpRtGxRAJr1b8aaMWswm8zsXb2XgQMHUqRIEQB8fHwACAsLo0ePHhQtmrHPoEGDeO655+45ZnJyMoMHDyYqKgqNRsP58+eJioqiSZMm92y7dOlS4uLiWLJkCUuXLgVAp9Nx7ty5bN+X/vhTjnK5eg0AaretjVcpLwCOrj/KjTM3+OGpHyyPKxqFWxduUeSJIiz/bDm/7/4dVLh27Rp+fn688MILAAQHB1OxYkUg45qJ9PR0SpcuDUCjRo04ceJErnMriObNm0dAQABGo5HXX3+dDz74gLp16wIZzWmyU7xycXybZnRqvHbiGhqdhurNqwNQvXl1SlTJ+LuYmpjKiU0niL8eb9k3LTGNayevWW537NXRsoZ+ixYteO+990hISKBly5aWgjgnSpYsSf369Vm7di29e/fOxSsgxMNJ8SwcRuXKlXnppZcsp5IhY0rGwoULLVM07ta2bVtWrFjBrVu3GPv22xzdsIEVR49y6uZNAsuWBeC3Hj2IunSJLdHRPDd7Np+2bk33Ox80/6YoCudjY3lv5UrCBwygSrFiHLpyhY53ndK/336qqlKnTh22bbt3Oa3o6Gjc3d2ztKINCAigRIkSrF27VopnKwkNDWXatGlUrFiRVq1aAdCyZUvCw8MJDw/PMjcUyDIKrdVqczSHU1VVPvroIwYMGHDPYzrlnz/X5f3Lc/30dRJjEvEo5pFtLJ3LP9trNJr7z0VV7v5nzi6IvV+DoI8//pgSJUoQGRmJTqfjmWeeISUl+xUsVFVFURR2796Ns7PzPY/9+315PO04qxJXZdmugn8Fts168BJzzh7/xFZVlZqtatL7l3uLqbVj1xJ/PZ5tO7bh6ebJu+++myX3f/88H+Xn+zjR6XR0796d9957j8DAQEwmE1pFi1E1kp6SnmXbu1/jbGX+utyZdTN07VCcXLPvKOjl6WX5d/fu3QkJCWHdunVMmDCB8ePHW6Y35YRer2fmzJmYzWZp1S2sSn6bhEP59NNP+f3337l0KeM0YteuXRk1apTlg+/WrVucPJnRuaxNmzasX7+e6OhoajVqRKuqVRkVEUGTihXRajQYTSbOxMQQWK4cbzVtSpfatdl7ZxoFwJw7KwrcSkpixdGjtKxShbjUVJy0Wkp7eqKqKr/s2mXZXnHPmLuaOT82OjqazZs307x5c0JCQjhz5gxhd60HHRUVdd9pGRqNhrZt28q8Zytq1KgR165dY/bs2VmK57lz53L58mXLaHRueHt7Z5k73LVrVyZPnkxMTMaqF+np6UTeWYXFR+tj2a5k1ZL4P+3PH2//QdLtjDnJqqqyf9l+bkTfeOAxqzerztH1R7l9OeO422Zso0aLGui0Olo91YrJkydbcoqNjcVkMtG2bVvmz59PXFwcAFOmTMFgMNwT+9atW5QvXx6dTsexY8eynDL/93Pt3LkzOp2O4cOHW+5bc+e6guzel7dO37rnePW71yc5Ltkycg4ZFxFunLIx2+deq3Utjm88nmUawdm9GXPKk2KTKFa6GJ5unly5csVyTYEj27BhAzVr1sTX15cDBw7gkeqByWhi38L7L2tYyrcUpnQTJ7dm/B09ufUkN05n/E66eLrg28yXsPH//B27ffk2sRdjgYxVX7w13pbHTpw4QenSpenduzejR49mx44ducrfYDBw9epVDh48mKv9hHgYGXkWDqVEiRK8/fbbjBgxAoBx48bx4YcfEhAQgEajQafTMXr0aHx9fSldujSlS5emYcOGaIoXp2m1alyJj+fNkBAATKrKG0uXEpucjFajoYSHBz916fLPsdzdaTllCnEpKQwICqLxnVO+z/j50WTiRIq5udGpVi3L9tpyGaekTSYTgYGBJCYm8sMPP1C5cmUA/vrrL4YPH86wYcNIT0+nYsWKLFmy5L7P1WAw0L9/f2JiYihWrJg1X0aH5OTkRLNmzdi/fz+17vzcatSoQXx8PM2aNcPJKfuRtAcZMGAAw4YNY9y4cXz99df07NmTmzdvWqaAGI1G+vXrR2BgIKW1pbNcMPfijy+yduxaxunHodVqMZvNVAupRvWW1bl14d5CM9MTtZ+g82edmfzcZAB8yvnQY3wPzJjp9XIvfG76EBISgk6nw8PDg7CwMDp06MChQ4cIDg5Go9FQr149Jk6ceE/sTz/9lJdffplZs2ZRrVq1LKuPdOvWjd9++81yweCkSZP49ddfeeONN/j5558BqF69Ou3atcv2ffnNqG/QlNZg5p+pA87uzry1/C2Wf76cLxt8iYuHCyjQ4NkG2T73klVL8vLPLzNv6DzSk9MxphkpX688vX/pTcuBLZnzyhzq1KlD2bJlczVFoDDp0aMHbm5uGI1GKlWqxOTJk6lQoQIdO3bkw+APcS3lSuXGlUncm31Ld52Ljj7T+vDne3+imlXKB5SnVPVSuBVxA+Dln19myadL+CbkGxRFwdndmefHPU/RckUBKKb952/Vn3/+ye+//46zszNms5nJkyfn6rk0bdoUNzc31q5di7+//6O9IEJkQ1Glk4IQOZI4dy7G48dztFxd3XHjmP3CC9R74okcx3ft0AHXxo25deuWZW5pXly4cIEKFSowf/78bOenisdLuprOz7E/Y8R20wP6eveliLaIzeLn1ZL4JZwznrN6k5RMrd1bU9cl+2lXAqLTo1masPSh26XEp+DqlTGt5dy+c0ztOZVP936a7RKCd/PR+NC7iHXnJ3fo0AGTySRn4YRVybQNIXLIuVEj263zrNPhXK+eVUOWL1+eJ598Uj40CgknxYk6LnVs1mGwkq5SgS6cAfxd/W1WOOvQUdO5pk1iFxaVdJXwVO6/wkum/cv3M7r5aEY1G8X8d+fTc1LPhxbOAAGuAVbIMiuDwcCmTZtITrZdcxfheGTahhA5pKtaFU3Jkphv3HhoEX1w6NCcB1YUnBs0QHF1tXpLbYPBwJIlSywXZ4nHm7+LPwdSD1g9ropKfdf6Vo9rbZV0lSiqKcpt822rFtEKCv4u/jgrDy/wHJmiKDR0bUhEcsQDt2v8UmMav9Q4V7FdFVdqOdd6+Ia5ZDAYePfdd9m8eXO28/SFeBQy8ixELixPS8uyfFieKQqKlxeuOehM9yj0ej1nz54tNMtoOTofrQ9NXO9d9i0vFBRqO9emolNFq8a1BY2ioZ1HO6sXzl4aLxq75a7Yc1R1XepSWlva6mdA2rq3tcmXl9q1a1O2bFlZ71lYlRTPQuRATEwM3bp1o+fQoUSkplovsKLg3q0birNtRrxatmyJk5OTTN0oRBq6NqSMtoxVihcFBU+NJy3cHp/lDMvoyhDkmvuVTe5HQaGdRzuclNxf8OmIMr/A6NBZrYB+0vlJqjlXs0qsf1MUBb1eL38DhVVJ8SzEQ2zZsoWAgAA2b97M0qVL6fa//+HcOI+jVIqSUTh3747uzmoatuDp6UnTpk1l1KUQ0Sgaunh2obi2eJ6KFwUFd8Wd7p7dcdG4WDFD22vi2oS6znm/sE+Dhk4enSirK2uFrByHj9aHrl5drVJAV3WqShv3NlbKLHsGg4EDBw5w5coVmx5HOA4pnoW4D5PJxJdffknLli2pXLky+/fvp3PnziiKgmu7dri2awdaLeR28X1FQfHwwKNXL5xq17ZN8nfR6/Vs2LCB9PT0h28sHguuGlee9XoWXyffR45RVleWF7xfKPAXCWZHURRC3UNp5tYMDZpcF3AKCh6KB908u1HVuaqNsizcyurK8pzXcxTVFM31vpk/r3K3yvFh8IcsXbzU6td73C1z2cG718kXIi9kqTohsnHp0iV69epFREQEI0aM4NNPP0Wnu/f6WtONGyT/9Rem6OiMItqcfftawDLa7BQYiFvbtih3dSizpT179tCoUSM2bdpE8+bN8+WYIv+cTDvJxqSNJKgJKCj3nQ+c+Zir4kqwWzB1nesWiotIb5pusiFpA5eMlx74/DNp0ODn7EdT96ZygaAVGFUju1N2sy9l30OXUcz8+ZTQlKC1R2t2LN/BM888A0DNmjX54IMPeOmll3Bxsf6ZkMDAQOrWrcuvv/5q9djC8UjxLMS/rFq1it69e+Ps7Jylm9yDmK5dI23vXoxnzty7GodOh7ZMGXQ1avDquHEcO3+eiIgIPDyyb6tsbSaTidKlSzN48GA+//zzfDmmyF9m1cxZ41mOpB7hkvESiWrWBhauiivF1eLU9aiLr5MvWkVrp0xt54bpBodSD3Eu/Ryx5tgsRbQOHcWUYlTVVaWue13cNe52zLRwSlVTOZp6lJPpJ7lqvEo6Wc90FdEUoZyuHH4ufhlz9hWF5ORkPD09Md816FCiRAmGDBnCG2+8YZX17jN98MEHzJo1i8uXLxeKL43CvqR4FuKOtLQ0Pv74Y7799ls6duzIzJkzKVmyZK7jqEYj5tu3wWRCcXJCKVIE5c7UjipVqmS0+65Vi9WrV1OpUiVrP41s9ejRg7Nnz+a6va14PCWbk0lWky0jzWm30yhWrBjPP/888+bNs3d6NmdUjcSb4zFjxgknvDRelClTBq1Wy6lTp3Bzc7N3ioWaqqokqAmkqWlo0OCh8bjvKH+DBg3Yt+/edt+DBg3Ktovlo1q/fj1t27Zl//791LPymvrC8cicZyGA06dP06xZM3744Qe+++47VqxY8UiFM4Ci06EtXhxtqVJofHwshTNgGUk5duwYAQEBbNy40RrpP5TBYGD37t3cunX/ts2i8HDTuFFMW4zi2uJ4aDwsxcn8+fOZNGmSnbOzPZ2iw0frQ3Ftcby13hiNRm7cuMHly5d58cUXrbvcpLiHomQs/1dcWxwfrc8Dp8e0aNEiS2t7jUZD3bp1+fDDD62aU9OmTXF1dZWLp4VVSPEsHN68efMICAggJiaGbdu2MXToUJud1ktJSQEyRmbi4uJo06YNEydOtOnFMpBx0aDZbGbDhg02PY4omPbs2WP5nX7jjTdYvHixnTPKX4cPH7ZMDVi6dClvvfWWzd9zImeCg4OzXMzs6urK8uXLqVjRuuuOu7q60rJlS1myTliFFM/CYSUlJfHaa6/xwgsv8NRTT7Fv3z4aNmxo82NmMpvNmEwm3njjDRYsWGDT41asWJGaNWvKB4eDurt4BnjhhRfYunWrHTPKX3v27Mlye9KkSfzvf/+zUzbibiEhIUDGaPV7772Hh4cHffr0scnqQJmtujMHMYR4VFI8C4d08OBBGjZsyJw5c5g+fTqzZ8/G29vb5sf99x9trVZLr169cnRRYl4ZDAbWrl0rI24OaOfOnZaRV1VVSU9Pp2PHjpw6dcrOmeWPvXv3ZpkaAPDJJ58wZ84cO2UkMpUvX54PPviAZcuWMXr0aBYtWsS2bdt45513rH4svV5PSkoKW7ZssXps4VikeBYORVVVpkyZQlBQEDqdjj179vDKK6/k29XXycnJAJYLliZNmsRvv/1GqVKlbH5sg8FAdHS0wxRMIkNMTAznz5+33FYUBVVVSU1NzXJ/YbZz5857RjIVReHYsWN2ykjc7ZtvvuGpp54CoFmzZvz0009MmjSJKVOmWPU4fn5+lClTRs7AiTyT4lk4jNjYWJ5//nkGDhxIv3792LlzJ08++WS+5vDuu+8yduxYrly5QsuWLZk+fXq+jQS3atVKWnU7oNOnT1v+XaZMGXQ6HWvWrOH27dv5csajIMh8DUqUKAHAiBEjuHHjBp999pk90xL38dprr/HGG2/w5ptvsmnTJqvFVRTFcgZOiLyQ4lk4hB07dhAQEEBYWBgLFy7kp59+sstyVSNHjmTYsGF4e3szfPhwduzYwbZt2/Ll2J6engQHB8sHh4MJCAhgw4YNXLt2jb/++ov09HRcXFxs0oiioAoLC+PYsWNcuXKFkiVLkp6eTrFixeydlniAcePG0bx5c7p3787Zs2etFlev17N//36uXr1qtZjC8UjxLAo1s9nMqFGjaNasGWXLliUyMtLS0creOnbsSK1atRg7dmy+HdNgMEirbgej0+kIDQ2lZMmSBAQEUKJECYf7AtWgQQNq1KiBVqtFr9c73PN/HDk5OTF//nw8PT3p0qULiYmJD98pB6RVt7AGKZ5FoXX16lXat2/PRx99xPvvv8/GjRupXLmyvdOy0Gg0DBs2jKVLl3L8+PF8OabBYCA+Pp5du3bly/FEwaLRaNDr9Q691q3BYGDfvn3cuHHD3qmIhyhRogTLli3j5MmTvPLKK1aZ4lamTBn8/f3lC5TIEymeRaG0bt06/P39OXDgAGvWrOHrr7++52r7gqBXr16ULFmScePG5cvx6tevj4+Pj3xwODC9Xs+ePXu4efOmvVOxi7Zt26KqKuvXr7d3KiIH6taty2+//caCBQv46quvrBLTYDCwbt06WXlIPDIpnkWhkp6ezkcffUS7du3w9/dn//796PV6e6d1X66urrz11lvMnDmT69ev2/x4Wq2Wtm3bSvHswPR6vUMXj+XKlaNOnTryHniMdOvWjc8++4z//Oc/LF26NM/x9Ho9ly9f5u+//7ZCdsIRSfEsCo3o6GhatGjB2LFj+eabb1i1ahWlS5e2d1oPNWjQIDQaDRMnTsyX4xkMBnbt2kVsbGy+HE8ULOXLl6d27doOP3VD1jx/vHz66ad0796dXr16cejQoTzFatasGa6urvIFSjwyKZ5FobBw4UICAgK4cuUKmzdv5v3330ejeTx+vYsXL06/fv2YMGFClg6EtiKtukXmRXOOWjzq9XouXLgg6zw/RjQaDTNnzqRq1ap06dIlT9OO3NzcaNGihRTP4pE9HtWFEPeRnJzMoEGDePbZZ9Hr9URGRtKkSRN7p5VrQ4YMISYmhl9//dXmx6pUqRI1atSQDw4HZjAYOHfuXL5dqFrQtGjRAmdnZ3kPPGY8PT1ZunQpcXFxPP/883laNUiv10urbvHIpHgWj60jR47QuHFjZs6cyZQpU5g/fz5Fixa1d1qPpFq1ajzzzDN89913mEwmmx8v84IZ4ZhatmyJk5OTw/4OeHh40KxZMymeH0OVK1fmzz//ZNOmTQwbNuyR4xgMBpKTk9m6dasVsxOOQopn8dhRVZXp06fToEEDTCYTu3fvZsCAAfnWYttWhg8fzokTJ1i+fLnNj6XX6zl9+rS06nZQHh4eNG3a1KGLR4PBQEREBGlpafZOReRSy5Yt+eGHH/jxxx+ZNm3aI8WoW7cupUuXduj3gHh0UjyLx0pcXBw9e/akf//+vPTSS+zevRs/Pz97p2UVjRs3plmzZvnSNKVVq1bodDr54HBgBoOB8PBwhy0e9Xo9iYmJbN++3d6piEcwaNAgXn/9dQYNGvRIo8eKojj8mufi0UnxLB4be/bsITAwkBUrVvDHH38wdepU3N3d7Z2WVQ0fPpytW7fa/APd29ub4OBg+eBwYAaDgYSEBHbs2GHvVOzCUbstFiY//PADTZo04ZlnnuH8+fO53t9gMBAZGcm1a9dskJ0ozKR4FgWe2Wzmu+++IyQkhOLFixMVFcULL7xg77Rs4umnn6ZGjRp8++23Nj+WXq9n/fr1GI1Gmx9LFDyBgYEUL17cYb9AZXZblOL58eXs7Myff/6Jq6srXbt2zfVqRdKqWzwqKZ5FgXb9+nWefvpphg0bxjvvvMOWLVuoWrWqvdOymcyW3YsWLeLkyZM2PZbBYCAuLk5adTsojUbj8A1z9Ho9e/fuddhui4VBqVKlWLp0KUePHqV///65Wn7xiSeeoG7dug77BVI8OimeRYEVHh6Ov78/u3btYuXKlYwZMwZnZ2d7p2VzL7/8MiVKlLB5y+6GDRtStGhR+eBwYAaDgd27dxMTE2PvVOzC0bstFhYBAQHMnDmTuXPnMmrUqFztKw1zxKOQ4lkUOEajkREjRtCmTRuefPJJ9u/fT4cOHeydVr5xc3PjzTffZMaMGdy4ccNmx9FqtbRp08ahRx4dXWbx6KgNczK7Lcp74PH33HPP8emnn/Lxxx+zYsWKHO9nMBi4dOkShw8ftmF2orCR4lkUKOfPnyc0NJSvvvqKL774grVr11K2bFl7p5XvBg8ejKqqTJo0yabHMRgM7Ny5k9u3b9v0OKJgqlChArVq1XLo4jFzxQUZeXz8ffbZZ3Tu3JmXXnqJI0eO5Gif5s2b4+LiImfgRK5I8SwKjKVLl+Lv7090dDQbN27kk08+QavV2jstuyhRogSvvPIKP/74o007YOn1ekwmE+Hh4TY7hijYHP20taN3WyxMNBoNv/32GxUrVqRz587cunXrofu4ubnRvHlzh/4CKXJPimdhdykpKbz99tt07dqVFi1asH//fpo1a2bvtOzu3Xff5caNG/z22282O0aVKlXw9fWVDw4HptfrOXv2rM0vUC2oMrstynugcPDy8mLp0qXExMTQo0ePHK0mlNkwJzU1NR8yFIWBFM/Cro4fP05wcDBTpkzhxx9/ZPHixRQrVszeaRUIvr6+dOvWjW+//Raz2Wyz42SOPArH1KpVK4cuHjNbdctp+8KjWrVqzJ8/nw0bNvD+++8/dHtp1S1yS4pnYTe//vor9evXJykpiZ07d/Lmm28+9i22rW348OEcO3YsVxfA5JbBYODUqVOcPn3aZscQBZenpychISEOWzxDxui7I3dbLIzatGnDuHHjGDduHLNmzXrgtnXr1qVUqVLyBUrkmBTPIt8lJCTQu3dv+vTpw7PPPsvevXsJCAiwd1oFUnBwMCEhITZt2R0aGopWq5UPDgeWWTymp6fbOxW7cPRui4XVm2++Sf/+/RkwYMADf7bSMEfklhTPIl9FRkZSv359Fi1axK+//srMmTPx9PS0d1oF2vDhw9m8eTM7d+60SXxvb2+aNGkiHxwOzGAwEB8fb7PfsYIus9uivAcKF0VR+Omnn2jYsCHdunXj4sWL993WYDCwb98+rl+/no8ZiseVFM8iX6iqyg8//ECTJk3w9PRk3759vPzyy/ZO67HQuXNnfH19bdqy22AwSKtuB1a/fn2KFSvmsMVjZrdFOftS+Li4uLBo0SJ0Oh3dunUjOTk52+0yW3VLwxyRE1I8C5u7efMmXbt25Z133mHgwIFs376dGjVq2Dutx4ZWq+Xdd99l4cKFNpuXbDAYuH37Nnv27LFJfFGwZTbMceTi0dG7LRZmpUuXZsmSJRw8eJABAwZkuyxj2bJl8fPzc9gvkCJ3pHgWVqOqKkOGDGH58uWW+zZv3kxAQABbtmxh6dKlfP/997i4uNgxy8dTnz59KFasGOPHj7dJ/IYNG1KkSBH54HBgBoOBXbt25Wht3MJIWnUXbg0aNGD69On8/vvv9z2L5+hrnouck+JZZKGqKqabNzGeP4/x/HlMN2/m+A/J6tWr+f777+nRowdHjx7liy++oFWrVlSpUoX9+/fTuXNnG2dfeLm7u/PGG28wbdo0bt68afX4Op1OWnUXEKqqcst0i8vGy1w2XuaW6Va+fJjr9XrMZnOBaNWdaE7kivEKl4yXuG68jlG1/XSizG6Ljjz6Xti9+OKLfPjhh3zwwQesWrXqnsf1ej0XL17k6NGjQDqQAqTe+bcU1OIfiipfsRyemp5O+t9/kxYVhenSJfj3Ffc6HdonnsDZ3x+nunVRnJ3vjaGq1K9fnwMHDqAoCq6uriQmJjJixAj+85//oNPp8unZFF7Xrl2jUqVKfPrpp3zyySdWjz9lyhTeeOMNYmJi8Pb2tnp8cX/pajrH045zOO0w14zXMJK1WNSho5SuFLWda1PDuQZOipNN8qhZsyahoaFMnjzZJvHvR1VVzhnPcSj1EBeNF0lWs85LVVAopimGr7Mvfi5+eGpsc5HxO++8w9KlSzlz5owsm1lImUwmunTpwpYtW9i5cyc1a9a884iZ1NRb7N4dQePGgTg5/bu7rQI4A56AOzL26NikeHZgqqqStmcPKevXQ2oqKAo87NfB2RnXVq1wbtwYRfPPH4+lS5fStWvXLJt2796dP//80waZO66BAweyZMkSoqOjcXV1tWrs06dPU61aNZYsWUKXLl2sGltkT1VV9qfuZ1vyNtLJ2TJxTjgR4haCv4u/1Qu8t956ixUrVnD69Ol8Kx4vpF8gLCmM2+bbKCioDxjhU8jIqbZzbZq7NcdFY90pYH/99RdPPfUUx44dk+syCrHbt2/TpEkTzGYzO3fuoGhRLRALqKiqmoPffQXwIaOQli9Zjki+Ojkoc3w8ibNmkbJyZUbhDA8vnAHS0khZu5bE6dMx376dEcts5qOPPrrnD87ChQuzzH8WeTd06FCuXbvG7NmzrR67atWqVKtWTaZu5JMEcwIL4hewMXljjgtngHTS2Zi8kQXxC0gwJ1g1J4PBQHR0NKdOnbJq3OyYVBMRiREsTFhInDkO4IGFc+bjKiqH0w7za9yvXEi/YNWcMlt1y9SNwq1IkSIsW7YMZ2cNV67sB26ROS0jZ18aVSAGuArICkWOSIpnB2S+fZuEadMwnTv3yDFMly6RMHUq5lu3mDhxIkeOHLHMy3Ry+ueU8vz58/Ocr/hHzZo16dy5s81adhsMBikc8kGcKY65cXO5YrryyDGumK4wL24ecaY4q+XVqlUrdDqdzX8HTKqJFQkr2J+2H3h40fxvKirJajKLExZzOs16K9BIt0XHUb16ZfbsWUH16hXzECUVuIIU0I5HimcHo6alkfjrr6hxcTkbab5vIBU1KYmEWbP4depUAIoXL067du0YMmQIM2fOZM+ePUy985iwnuHDh3PkyJFsL3jJK71ez4kTJzhz5ozVY4sMaWoaCxMWkqQm5bpovJuKSqKayMKEhaSp1mkr7eXlRXBwsM2Lx/VJ64k2RucphoqKGTN/Jf7FFeOjfwn5N4PB4NDdFh2DGbiKi4sOrfbfc5tzy0TGCLT1BzNEwSVznh1M8qpVpO3enbfC+W6KguLnh7FVK4oVK2admOKBVFUlJCQEV1dXwsPDrRo7NjaWEiVKMHHiRAYMGGDV2CLD+sT1/J32d54K57spKPg5+9Hao7VV4n355ZeMGTOGmzdv2uRC31Npp1iRuMJq8RQUvDXe9PLuhU7Je7579uyhUaNGbN68mWbNmlkhQ1Hw3AASrRzTC5DPQEchI88OxHjxImm7dlmvcIaMEeiDB/GOs96pY/FgiqIwfPhwIiIirN7UpGjRojRu3FimbtjIJeMlDqUdslrhDBkjsAfTDnLJeMkq8fR6PXFxcezatcsq8e6WrqYTlhRm1ZgqKrfNt9mTYp33QmBgoEN3Wyz8UrB+4QwQT8Y0DuEIpHh2IKk7doDGBj9yjYbU7dutH1fcV9euXalatSpjx461emy9Xk9YWBgmk8nqsR1dZEqkZcUIa1JQiEyJtEqshg0bUrRoUZsUj8fSjpGiplg9LkBUapRV1oPWarW0bdtWiudCy5YDPTKI5CikeM6DgIAAAgICqF27Nlqt1nK7R48eREREEBAQkOdj9O3bF2dn5yxtmYcPH87IkSNzFcecmIjx8GF4wEVmX6xfz6v/Wlpu2IoVvPfXXw8JbsZ4/Djm27dz9Lz37Nljea0qVqxIkSJFLLfHjBlDx44dOXbsWE6fmkPKbNm9YMECoqOjrRrbYDAQGxtbqFt1V65cmZo1a+Lv74+vry9dunRh27ZtjxxPURRiY2MfuE2iOZHjycdZNWoVXzf+mm9CvmF0i9HMGzKPpNtJnNhygtEtRj/S8VVUTqWfItGcsxG1kSNHMmTIkGwfu7t4jIiIYPXq1Y+UE4DRaOSzzz6jVq1a+Pn50alhJ8vztYbLhy/zmf9nAKSqqZxMP2mVuJmtuh2122JhsGjRIho0aEBAQAC1atWidevWmM1pQPJD9310SWTMgRaFnXSuyIOoqCgAoqOjCQgIsNwGiIiIsNpxypUrxyeffMIff/zxyDFM5849sHAG+KBlS1r/8gvLDh+mc+3abDx9mo1nzrDp9ddzdIyUHC5v1bBhQ8trNXPmTJYsWcKSJUssj7/33ns5iuPo+vbty4gRIxg/frxV23YHBQXh7e3NunXraNy4sdXiFjTz5s2zfNFbtGgRHTt2ZM2aNbl6zkajMcfzgi8aL/LH23+QdCuJIWuG4F7UPWOd56X7SbqV92JSReWi8SI1nPO+PrFer2fw4MGsXr2alJQU2rdv/0hx+vfvT0xMDNu3b8etiBuTYydbnq97Efc853k3BYXz6eep5Vwrz7Hu7rbYvXt3K2Qn8tPly5cZMGAAe/fupVKlSgDs27cPRbHttIqMvwepZDRREYWZjDzbkNFoZPDgwfj7+1OnTp0sI3lr1qyhWbNmNGjQgKCgoAde+DVgwAC2bt3Kvn37sn187NixBAUFUb9+fdq3b8/Zs2eBjKL70qWMeZAvDByIYdo0AFKNRqqMGkWqMespTmedjsnduvHx6tVEx8TwzvLlTOzalcS0NHrNnUvIxIkE//QTM+56HnXHjeO/69bR+pdf6DdsWJZ4cXFxGAwGPv/881y8ahmjgpnFdatWrRg2bBgtWrSgYsWK/Oc//2HlypU0a9aMypUr891331n2O3HiBJ06daJRo0bUq1ePCRMmAJCcnEyPHj2oXbs2/v7+GAyGXOVTUHl4eDB48GCmTp1q1REynU5H69atHeq09TPPPMPAgQMt02DWr19PcHAwgYGB1KlTh2l33juQ8aWlX79+tGjRAj8/vyxxVFXlgw8+oHPnziQlZS2I9x3bR9TSKF6c8CLuRTM+XBVFIaBrACUqlwDAbDSzYPgCRjcfzTfB33Au8p/lJHfP282oZqMY1WwUP/f4mdhLsZbHwr4PY1TTUXRs0JEmTZpYjj1mzBjq1KlD3bp16dmzJ7fvrM1+t4MHD9KsWTPq169P7dq1+fLLL9Hr9ZhMJqZMmcLs2bMJCAiwvI9z+rfr5MmTLFiwgBkzZuDj48M10zXL83V2c2ZC5wmMDR3LN8Hf8Of7f1qWXtw5ZycTu01k1quzGNV0FN+2/pYb0TcscVd9s4ovG37J2NCx7Fv8z9/EBe8v4OdRP1tuHzt2jAoVKmA05n4qR8WKFalZs6ZDvQcKk6tXr6LVarNcxF6/fn0UJZ3hw7+iUaPOBAR0oEWL5zl2LGPQ56uvJvDmmyMs2yckJFKsmD/Xr9/EZDLx3ntf4+dnwM/PwFtv/Ze0tIwVbvr2HUa/fu/RosXz+PkZGDt2bJaLrTMvwo6JicmnZy/ygxTPNnT06FH69OnD/v37eeuttywtlU+fPs3IkSNZuXIle/fuZc6cObz00kukpmb/rdjNzY3//ve/fPDBB/c8NmfOHI4dO8b27dvZt28fPXv2ZPDgwQC0adOGsLAwzGYzB0+cIC4lhbiUFHacO0fAE0/gks2ImV+ZMvRv1IgWU6bQrU4dgipU4P1Vq6heogTbBg9mWZ8+jN20id3nz1v2iUlKYv2rrzLt1Vct950/f57Q0FB69erFiBEj7jlObpw9e5bw8HD279/PDz/8wMqVK9m8eTNbt25lxIgRxMbGYjKZePHFF/n222/ZvXs3O3bs4Oeff2b37t2sXr2a2NhYDh8+zP79+5k7d26e8ilI3nzzTYxGI1OmTLFqXIPBwPbt24mPj7dq3IKscePG/P3330DGB+2WLVuIjIxk8+bNfP7551y48E9Djr179/LXX39x9OhRy32pqam8+OKLJCQksHjxYtzds44+RUZGUrJqSTyL37+19LUT12j0QiPe3/w+zQc0568vM6ZMXT58mWX/Xcbr81/ngy0fUDmoMnPfyfg93vXHLg4sP8Dbq95m3PZxrFq1ChcXF1atWsX06dPZunUrBw8exMPDgw8//PCeY1auXJn169ezb98+9u7dy8KFC7l69SrVq1enWrVq9OzZk6ioKEaMGJGrv1379u2jevXqlCiR8cUg1hxrecytiBuvzXmN4eHDeX/L+8SciyFqcZTl8XP7zvHUf57ig60fUKNlDdZ/vx6Av9f+zf6l+xm+YTjDNgwj5tw/BUnz15qzYeYGy1z9zBVjHnXFEIPBwNq1a5EFqR4/9erVo1mzZlSqVIlu3boxZswYLl68CKTzwQcD2b17GVFRqxg8uBfvvJMx7ad372eYP/8vy+/yggV/ERoaTMmSxfn55zns3r2fvXuXExW1klOnzjJu3D9fqPfuPchff03n6NENvPrqSyxZssQyjWvGjBl06dJFVqMqZKR4tiFfX1/LKeDg4GBL167Vq1dz8uRJWrRoQUBAAM8++ywajYZzD2ha0rdvXy5evHjPKghLliwhLCzMMrdr9OjRljht27YlLCyMyMhI/CpWpHmVKmyJjibi9GlaVq1632O93bQpcampvBUSAkDE6dP0bdAAgJKenjz15JNE3DUH+6WAABRFQb2zLurVq1dp0aIFo0ePpnfv3rl92e7x7LPPotVq8fHxoWrVqjz11FMoikK5cuUoWbIk0dHRHDt2jL///psXXniBgIAAQkJCiI+P5/Dhw/j7+3PkyBEGDx7MvHnzsjRxedyVLl2a3r1788MPP9z3y9ejMBgMGI1Gqy+FV5DdXSTdvHmT5557Dj8/P1q3bs3Nmzc5dOiQ5fHnnnsOLy+vLPt36tSJOnXq8NNPP2W7dqxJffhcyBJVS1C5YWUAKjeqzM3omwCc2HKCWm1qUbRsUQCa9W/Gic0nMJvM/L3mb0L6huDm7YZRNeLj44NWqyUsLIwePXpQtGjGPoMGDcp2FZXk5GReffVV6tatS5MmTTh79ixRUVEYDIZ7Og0+yt+uu59/5sWSqlll+WfLGd18NGNajuF81HkuHrpo2bZyo8oUr1T8ntfh+MbjBHQNwNXbFUVRCOkbYtmndPXSlK5ZmiVLlpCYmMgff/yRp+UW9Xp9vnVbFNal0WhYuHAh27Zto3379mzdupU6depw8uRp1q3bQnBwN/z8DHz++Q9ERR0GoEKFsgQG1mbZsozVYGbO/JNXXnkOgLCwrfTt+ywuLi7odDpee+0F1q3bYjnec891wssr40tx0aLePPvss0yfPh1VVZk0aRJvvvlmPr8CwtZkzrMNubq6Wv6t1Wotpw9VVUWv1zNnzpwcx9JqtXz99dd8+OGHtGrVynK/qqp89NFH2X5ItG3blo8++ojatWvT2t+fYvHxRJw6xe4LF/iuU6f7H+vOihya+7Qp/fe9Hs7OGfffKUqLFi2Kr68vK1asoHXr1jlsd3p//34ds3tdnZycKFasWJZ553c7fPgwGzZsICwsjPfff5+oqCh8fHzylFdB8e677/LLL78wZ84cXnnlFavErFatGlWqVGHdunV07tzZKjELut27d1umYQwcOJCOHTuycOFCFEWhfv36pKT8s0qEp+e9o8etW7dm3bp1vPPOO3h7e9/zePWA6lw/fZ3EmEQ8inlkm4PO5Z8/yRqNBpPxPgX3fd5STsr9vxje73348ccfU6JECSIjI9HpdDzzzDOkpKSg1+v56aefskz1yM3frvr163PixAlu3rxJ8eLF0SpayxJ9ERMjiL8ez9B1Q3FydWLxJ4tJT/mnKYnO9a7XQXv/1+Hfz6nV66346aefuHHjBnq9ntKlSz80z/vJ7La4du1afH19HzmOsJ9atWpRq1YtXn/9ddq3b8+cOYsZP/5ndu9eRrVqlThw4AgtWjxv2b5fv+eZMWMBDRrU5eTJs7Rv3zLbuP/+vfP0vPssk8Lbb79N586defLJJylZsiSBgYG2eHrCjmTk2Q7atWtHWFgYBw4csNyXkzVVu3btiouLC4sWLcpy3+TJky3zqdLT04mMzFiyqmzZshQpUoTJkyfTpnlzmlepwprjxzkXG4v/E0/kON9WVavy65351jcSE1lx5Aih1apl3UijQXPn9GxmjpcuXeK1116zSRvpf6tZsybe3t7MmDHDct/JkyeJiYnhwoULKIpC586dGTt2LKqqcv6uaSePu1q1avH0009bnpu1ZJ62dgRLly5l0qRJDLszb//WrVtUqlQJRVHYtGkT+/fvf2iMjz/+mGeeeYa2bdty8+bNex6vU70O/k/7Z1w0eGe1CVVV2b9sf5Y5vdmp3qw6R9cf5fbljEJ224xt1GhRA41Wg18HP7bN3EZqXCpFNUUt05jatm3L/PnzibuzBvuUKVOyne9/69Ytypcvj06n49ixY5bR6dDQUBRFybLyTW7+dvn6+tK9e3f69+9PbGwsPhofy/M9v/883qW9cXJ1Iu5qHPuXPvz1BajZqiZRS6NIiU9BVVW2zcq6QkqTtk24cuUKX375ZZ5H+7y8vKRV92Pq4sWLbN261XL71q1bnDlzhiJFfHBycuKJJ0qhqioTJvyaZb+uXQ3s3n2A//1vIr16dbVM+Wnbtim//rqItLQ0jEYjU6fOw2Bofp+jO1GrVi2qVq3KgAEDZNS5kJKRZzvw9fVlzpw5vP766yQlJZGWlkZgYGCORnNGjRpFixYtLLd79uzJzZs3CQ0NBTIuUuzXr5/lm27btm1ZsWIFvg0bknTqFKU9Pann5YUmF+s9j+rQgXdXrCBk4kRUVWVYixY0LF8+60ZmM9onnoA7RbyTkxNz5szh1VdfpWfPnvz222826VaWSafTsWLFCoYMGcK4ceMwmUyUKFGCOXPmcPDgQT766CNUVcVoNPLyyy9Tr149m+ViD8OHD6dly5asXr2aDh06WCWmwWBgypQpnD171nLFemHSo0cPXF1dSUxMpHbt2qxcudIyzeqbb75h8ODBfPHFFwQEBOR4BY4hQ4bg4eFB69atWbNmDWXKlLE8VkpXihd/fJG1Y9cyTj8OrVaL2WymWkg1qreszq0L97/o84naT9D5s85Mfm4yAD7lfOgxvgcAjXo04vaV24xrN47pztMp5lmMsLAwOnTowKFDhwgODkaj0VCvXj0mTpx4T+xPP/2Ul19+mVmzZlGtWjVat87oVOjt7U2DBg04fPgwAQEBPPPMM4wYMSJXf7umT5/Ol19+SePGjdHqtNxMv0m1kGo8879nmNF3Bt8Ef4N3GW9qtMzZCiG19bU5u/csY0PH4urlypNtn7Q8pkHDE05P0L9/f+bMmUNwcHCOYj6IwWBg9OjRpKenF6rpXoWd0Wjk888/58yZM7i7u2M0GunTpw/vvPMWp08fp04dPcWL+9C1a9Yvky4uLjz/fCcmTvyNI0f+aeYzYMBLnDp1jvr1nwKgVasmDBnS7z5HzzgT+9prr/Hmm2/y7LPP2uQ5CvuS9twOwpyURPy33z50ubpHpih4DR2K5l/zQEX+UFWVxo0b4+Xlxfr1660SMzY2luLFizNlyhRevetiUPFoksxJTL091ardBe+moPBqkVdx11hvmazPP/+c7777jhs3bljly+8fcX9w3XTdZq9BB48OvPvMu/To0YOXX345z/F27dpF48aN2bJlC02bNrVChsK+jMDFh26VN+UBLW+++SalS5fmP//5j42PJ+xBpm04CI27O05+fjbrMKirWVMKZzvKbNm9YcOG+y5pmFtFixYlKChITltbibvGHV8nX5t1GKzuVN2qhTNkjLzevn2b3bt3WyWev4u/zQrnq1FX6Vi7IxqNhpdeeskqMRs0aICPj4+8BwoNHeBmw/geXLp0lVq1arFv3777NiISjz8pnh2IS5Mmthl5NptxscIpUpE3zzzzDFWqVOHbb7+1WkyDwSCtuq2ovmt9mxSPKiqBrta/KMnarbprONfATbFN8dK1SVdOnjzJsmXLsl3t5FFkdlvMbpUS8bi692Jea8YuW7YsR48eZdu2bfesyCMKDymeHYj2iSdwDg6GPK5+kYWi4NSgAbqKFa0XUzwSnU7H0KFDmTdvnqVRTl4ZDAZu3bpltdFsR1dGVwZ/F3+rjj4rKAS4BFBGV+bhG+dSZsMcaxWPOkWH3kNvlViZFBSKaYrRwLWBVeNm0uv17Ny586Ht18XjISYmiTVrttlgQMCbzPnOovCT4tnBuIaGoile3DrTNzQalKJFcdNb98NQPLpXXnkFb29vvv/+e6vECwoKwsvLS05bW1FTt6YU0RSxSgGtoFBEU4QQt5CHb/yIDAYDO3bsyLY74aOo4lQFP2e/h2+YQxo0tPNoh1axzmjzv93dqls83rZu3UpAQACvvfYeKSm57zx5f05AESvGEwWdFM8ORnFywqN3bzRFiuRtBFpRUDw98ezdG8XFxXoJijzx9PRk0KBB/PLLL1YZKXNycnK4Vt225qQ48YzXM3hpvPJUQCsoeGm86O7V/YHrO+eVwWDAZDJZtWFOqHso1ZyqPXzDB1BQ0KDhac+nKaUrZaXM7lW5cmVq1Kgh74HHmMlk4quvvqJly5ZUqlSJrVu34+FRFbDGFy4dUAoppxyL/LQdkMbLC4/+/dFWqfLIMbQVK+L56qto7nQvEwXHm2++SVpaGj///LNV4jliq25b89J48bzX85TTlXvkGOW05Zjx/Ax++f4X0tPTH77DI6pSpQrVqlWz6rxfjaKho0dHGrhkTLXI7ZcIBQUPxYPuXt2p5GT7ZRT1er3Me35MXb58GYPBwH/+8x8+/vhjwsPDqVChAhlFbxkgL4M/rndiyKq/jkaWqnNgqqqSHhVFyrp1qMnJGSPR9/t1yHzM1RXX1q1xbtgwz50Dhe3079+f1atXc+bMGZyd8zYP78SJE9SoUYPly5fz1FNPWSlDARnvwUNph9iavJVUNRUF5b4XFGY+5qq4EuIWgq/ZF3f3jNU1SpcuzfDhw3nttdcoUsT6p48HDx7MunXrOHHihNVjXzZeJiwxjBhzzAOffyYNGorfLE6bYm0oXezROwjmxrJly+jSpQsnT56k2r8bRIkCa/Xq1fTu3RudTsfs2bMt/RCyUoEEIBbI6QX1GsAH8OC+7T5FoSbFs0A1mUg/coT0/fsxXrgAd7UhBsDFBW25cjj7++NUuzaKDZudCOs4fPgwderUYdasWfTu3TtPsVRVpWrVqjz99NP88MMPVspQ3M2oGjmVforDqYe5arpKqpqa5XEXxYXS2tLUdqlNNadq6JSM92Dt2rU5cuQIkLFcoZubGwMGDGDo0KFUtOJFvEuWLKFbt26cPn2aKnk4Y3U/qqpyyXSJQymHuGC8QIKakOVxLVpKaEtQzbkadZzrUNS1KE5OTqxatSpL0yhbiYuLo3jx4vzwww8MGjTI5scTeZOWlsYnn3zC2LFj6dChA7NmzaJkyZIP2UsFEu/8n3rn9t00mExatNoigDtSNDs2KZ5FFqqqosbHoyYlgaqiuLujeHvLKPNj6KmnnuLcuXPs378/zz+/119/nU2bNlkKNWE7qqqSoCaQbE4GwE3jhqfime3PcMiQIUyYMOGelQPq1q2bpYV2Xt2+fZvixYszceJEBgwYYLW495NiTiHeHI8ZM86KM0U0RdAo/8wy9Pb2Jj4+HkVRGDZsGF9++SUuNr72onnz5pQsWZJFixbZ9Dgib06fPs2LL75IZGQk33zzDUOGDMlVR90MKmDin5FoDX//fQw/Pz8++ugjvv76a+smLR47MudZZKEoChpvb7RlyqB94gk0RYpI4fyYGj58OAcPHrTKhU56vZ6jR49y7tw5K2QmHkRRMi4ELKUrRSldqYwLC+/zHgwJCclSOGs0Gry9vfnuu++smlORIkVo3Lhxvl0056pxpaSuJKV1pfHR+mQpnCGjgQ9kfNH49ttvqV+/PgcPHrRpTgaDgfXr12M0WnOVBmFN8+bNIzAwkBs3brB161befffdRyicIWNUWUfG0nPOgI7IyEgA/ve//7Fs2TLrJS0eS1I8C1FItWzZkgYNGjB27Ng8x2rdujUajUYumipggu9qTqQoCqqqMmvWLNq2bWv1Y2UWjwWhYU7mXG/IKKCPHTtG/fr1Wb16tc2OaTAYiIuLY9euXTY7hng0SUlJDBgwgBdeeIEOHTqwb98+GjVqZNVjREZGWr7EPv/882zfvt2q8cXjRYpnIQqpzJbdYWFhREVF5SlWsWLFaNSokRTPBUyFChUoXTrjorl27doRGBjIW2+9xZUrV6x+LIPBQGxsLHv27LF67Ny6u3iGjAK6SJEiNu3oltltUd4DBcuhQ4do1KgRv//+O1OnTuWPP/6wyUWzO3fuJHOWa3p6Oh06dODo0aNWP454PEjxLEQh9uyzz1KxYkWrtOzOXK6rIIw8in+MGDGCr7/+mr/++otly5ZhMpno3r07qampD985Fxo1akSRIkUKxHrHHh4eAJZT8t27dyc6OpqmTZva7JharZY2bdoUiOcvMr4w/fzzzzRq1AiNRsOePXvo37+/TaYZms1my7SNzNvx8fG0bduWGzduWP14ouCT4lmIQiyzZffcuXM5f/58nmIZDAZiYmKyfIgI+xs8eDAfffQRGo2GcuXKsWjRIvbs2cMbb7yBNa8Hz2zVXRCKxxIlSqDVann99ddp06YN+/bty/OSjDlhMBikVXcBEBsbS48ePXj99dfp27cvu3btonbt2jY73okTJ0hKSrLcVhQFs9lMTEwMFy5csNlxRcElxbMQhVz//v3x8PDI8zJzTZo0wdPTU05bF3BNmjTh559/Ztq0aUyYMMGqsTMb5sTFxVk1bm5NmjSJU6dOMXHiRMaNG8fp06et1hToQfR6vdW7LYrc2blzJ4GBgaxdu5YFCxYwadIk3NzcbHrM48ePW/5dvHhxvLy8CAsL49atWwQEBNj02KJgkuJZiELOy8uLQYMGMWXKFG7fvv3IcaRV9+OjT58+DB06lKFDh7J+/Xqrxc0sHiMiIqwW81GUKVOGSpUyOgvWrVuXvn378tlnn9m8qK9SpQq+vr7yBdIOzGYzo0ePplmzZpQpU4aoqCieffbZfDl2mzZtWLlyJdeuXWPmzJnExcVRsWJFmy+PKAouKZ6FcABvvfUWKSkp/PLLL3mKo9fr2bp1KwkJCQ/fWNjV6NGjad26Nc8//zynT5+2Ssxq1apRtWrVAvcF6vPPPychIYExY8bY/FgGg6HAPf/C7urVq3Ts2JEPPviA4cOHs2nTJipXrpxvx3d3d6dDhw6ULFmSVq1a4eTkJL8DDk6KZyEcQNmyZenZsyfjx48nLS3tkeMYDAbS09PZuHGjFbMTtqDT6Zg3bx7FihWjc+fOxMfHWyVuQSwey5cvzzvvvMN3333HpUuXbHosg8HAqVOnrPaFRDxYWFgYAQEBREZGsmbNGv73v//h5ORkt3w8PT0JCQkpcO8Bkb+keBbCQQwbNoyLFy8yf/78R45RvXp1KlWqJKetHxM+Pj4sW7aMc+fO8fLLL2M2mx++00Po9XpOnDhBdHR03hO0og8//BBXV1dGjhxp0+O0atUKrVYr7wEbS09P55NPPsFgMODn58f+/fsxGAz2TgvIeA+Eh4eTnp5u71SEnUjxLISD8PPzo3379owdO/aRV2FQFAW9Xi+jLo+RJ598kj/++INly5ZZpbAsqA1zihYtyqeffsq0adNs2ka+SJEiNGnSRN4DNnT27FlatWrFqFGj+Prrr1mzZg1lypSxd1oWBoOB+Ph4du7cae9UhJ1I8SyEAxk+fDj79+/P00VkBoOBI0eOyBJNj5FOnTrx9ddf88UXX7BgwYI8xSpatGi+turOjcGDB1OxYkU+/PBDmx5HWnXbzqJFiwgICODixYts3ryZDz/88BFbbNtO/fr1KVasWIF8D4j8UbB+I4UQNtW6dWsCAgLy1LK7TZs2KIpS4EYexYN98MEHvPjii/Tt2zfPHSf1en2BadV9NxcXF7766iuWLVvGli1bbHYcvV7P7du3C0S3xcIiJSWFN954g+7du9OmTRsiIyOztJ8vSDIb5sjfQMclxbMQDkRRFN577z3WrFnDgQMHHilGsWLFaNiwoYy6PGYURWHq1KnUqlWLLl26cO3atUeOZTAYuHXrFnv37rVihtbxwgsvUL9+fd577z2rNom5W0HqtlgYHDlyhMaNGzNt2jQmTZrEggUL8PHxsXdaD2QwGNi1axe3bt2ydyrCDqR4FsLBPPfcc1SoUCFPLbsNBgNhYWFWuQBN5B93d3eWLFlCamoqzz777COvvBIUFIS3t3eBLB41Gg2jR49mx44dLFq0yCbH0Ol00qrbClRVZcaMGTRs2JD09HR27drFwIEDbdJi29r0ej1ms5kNGzbYOxVhB1I8C+FgnJycGDJkCHPmzHnkecsGg4EbN27k+fS/yH8VKlRg4cKF7Nixg7fffvuRYmQ2zCmop63btGlDu3bt+Oijj2y2IoLBYGDHjh1277b4uIqPj+fll1+mX79+vPjii+zevZt69erZO60cq1SpEjVq1Ciw7wFhW1I8C+GAXn31Vdzd3R+5ZXeTJk3w8PCQkbfHVNOmTZk0aRJTpkxh0qRJjxRDr9ezbds2q60fbW2jRo3i5MmTTJ061SbxM7styshj7u3du5f69euzbNky5syZw9SpU/Hw8LB3WrlmMBhYs2aNzaYHiYJLimchHJC3tzevv/46U6ZMeaSRM2dnZ0JDQ6V4foz179+ft956i7fffvuRmt4YDAaMRqPdW3Xfj7+/Py+//DIjR460SYFftWpVqlWrJiOPuaCqKuPHjyc4OJiiRYsSGRnJiy++aO+0HpnBYCA6OppTp07ZOxWRz6R4FsJBvf322yQlJTFt2rRH2t9gMLB161YSExOtnJnIL9999x0tW7bk2WefzXXTk2rVqlGlSpUC/QXqiy++4Pbt23ma3/8gBbHbYkF148YNOnfuzNChQ3nrrbfYunUr1apVs3daedKqVSt0Op18gXJAUjwL4aDKly/PSy+9xPjx4x9pXqjBYCAtLY1NmzbZIDuRHzJbeHt7e9OlSxcSEhJyvG9mw5yCXDhUrFiRt99+m7Fjx3LlyhWrx9fr9Zw8eZIzZ85YPXZhsnHjRvz9/dmxYwcrVqzg22+/xdnZ2d5p5ZmXlxfBwcHyBcoBSfEshAMbNmwY586d488//8z1vjVq1KBChQrywfGYK168OMuWLeP06dP07ds3VyuoGAwGjh07xtmzZ22YYd589NFHODs789lnn1k9dmhoqLTqfgCj0cjIkSNp3bo1NWrUICoqik6dOtk7LasyGAxs2LBBGuY4GCmehXBg9erVw2AwMGbMmFxf9KIoipy2LiTq1KnD77//zsKFC/nyyy9zvF9BbdV9Nx8fHz755BN++eUXjh07ZtXYBbnbor1duHCBNm3a8MUXXzBy5EjCwsIoV66cvdOyOoPBQFxcHLt27bJ3KiIfSfEshIMbPnw4kZGRhIeH53pfg8HA4cOHuXjxog0yE/mpS5cufPHFF/z3v/9l8eLFOdrHx8eHRo0aFejiGeCNN96gXLlyfPTRR1aPXVC7LdrT8uXL8ff3/3979x0W1bX1cfw7QwcRK5ZYsWCwAYq9ojOxxY7daOyxx27sGns0mmgsscTeexLF3guigNhRLLFXFKTPnPcPI2+8NsCBQ1mfPPe5kXPOnt8QYdbs2WcvgoODOXjwIKNGjcLMzEztWEmiTJkyZM6cWd5ApTNSPAuRztWuXZtSpUolqmW3tOpOW0aMGIGXlxft27cnMDAwXte8aZiTkotHa2trJk6cyJYtWzh+/LhJx9br9YSEhEirbiAqKor+/fvTsGFDqlSpgr+/P1WrVlU7VpJ606pbiuf0RYpnIdI5jUbDoEGD2LlzJ+fPn0/QtVmzZqVMmTLywpFGaDQali5dSpEiRWjYsCFPnjz55DV6vZ5nz55x9uzZZEiYeG3atMHV1dXkbbtTcrfF5BQUFETFihWZN28es2fPZuvWrWTNmlXtWMlCr9dz6tQpQkJC1I4ikokUz0IIWrZsyRdffMHMmTMTfK206k5b7Ozs2Lp1K69evcLLy+uTO7GUL18ee3v7FP/pg1arZerUqRw/fpxt27aZbFxzc/MU3W0xOaxcuRJ3d3fCwsI4ceIEffv2TRUttk3lTavuxCx9E6mTFM9CCCwtLenfvz8rV67k/v37CbpWp9Px+PFjAgICkiidSG758+dn06ZNHD16lO+///6j51pYWKSahjl6vR6dTsewYcNMujuCXq/nxIkT6a5Vd1hYGB07dqR9+/Y0adIkrnNgelOgQAGKFCmSKn4GhGlI8SyEAKBr165YW1vz66+/Jui6ihUrSqvuNKhq1arMmTOHuXPn8vvvv3/0XL1ez/HjxxO0T7Rapk6dypUrVxLdHOh9Unq3xaTg7+9PmTJl2LhxI8uWLWP58uXY29urHUs1svNQ+iLFsxACAAcHB7p168a8efMS1M7YysqKGjVqpOuPrdOq7t27891339GrVy+OHj36wfP0ej0xMTGJavOd3Nzc3GjXrh1jx441WXfMQoUK4eTklC5+BhRFYc6cOVSoUAFbW1vOnDnDN998o3Ys1el0OoKDg6VVdzohxbMQIk6/fv0ICwtjyZIlCbpOp9Nx5MgRwsPDkyiZUMvs2bOpVKkSTZs25fbt2+89p3DhwuTPnz/VzLxNmDCBZ8+eJWqN/4fodLpU8/wT69mzZzRt2pQ+ffrQrVs3Tpw4gbOzs9qxUgRpmJO+SPEshIiTN29eWrZsyc8//5ygNaHSqjvtsrCwYMOGDdja2tK4ceP3vkFKbQ1zChQoQO/evZk2bRqPHj0yyZh6vZ6rV69y8+ZNk4yX0hw9ehRXV1cOHTrE1q1b+eWXX7C2tlY7VoqRMWNGadWdjkjxLIR4y6BBg7h16xabNm2K9zXFihUjT548MuuSRmXPnp1t27Zx5coVOnXq9N6t3vR6PZcvX+aff/5RIWHCjRgxAnNzc8aPH2+S8VJDt8X4CA4OpnPnzjx//hwAg8HAjz/+SPXq1cmfPz/+/v40atRI5ZQpk06nk1bd6YQUz2mMoigYQ0Iw3L+P4f59jC9fmnRPU5H2ubq6Urt27QS17NZoNOniY+v0rHTp0ixfvpx169YxefLkd457enrGNcyJMEbwOPYxD2Mf8tzwHKOS8rYxzJIlC8OHD2fBggUEBQV99niZMmWiXLlyKaJ4NigGnhme8TD2IY9jHxOlRMX72iFDhrBkyRI6d+7M3bt30ev1jB49mh9++IEDBw6QL1++JEyeuun1el68eMHp06fVjiKSmEaRyirVU2JiiLlwgehz5zDcvQvR0W+fYG2Ned68WJYujXmxYmjSaJtUYTre3t7UqVOHgwcPUr169Xhds3btWlq3bs3du3fJnTt3EicUahkzZgwTJkxg27ZtfP3118DrN+3/xP7DrD9nkcstF5YOlm9dY4YZ2c2yU9SyKF9afom1NmV83B8REYGzszPly5dnw4YNnz3emDFj+PXXX3n8+HGyt6MONYZyPuo8wTHBPDM8w8jbb1gyajOSzzwfJaxKkMM8x3vHCAgIwNXVNe7PGTJkwN7enpUrV+Lp6ZmU8dOE2NhYsmfPTv/+/RkzZozacUQSkuI5FVMUhWgfHyIPHICoKNBo4EP/Of89prG1xVqvx6JUqXS1ib1IGEVRKFWqFAUKFGDHjh3xuubJkyc4Ojryxx9/yN33aZjRaKR58+bs3buXkydPkqFIBvaH7+eF8QUowCd+rZhhhquVKxVsKmCuMU+WzB+zbNkyOnbsyMmTJylfvvxnjXXs2DGqVKnCqVOnKFeunIkSflyEMYLD4Ye5EnMFAIUPv6Rr0KCgkNMsJ7XsapHNLNtbxxs2bMjff/8d12pdo9Gwb98+atasmXRPII1p1qwZDx8+/OjuNCL1k+I5lTK+eEH4xo0Y7txJ1PXmhQtj06QJWltbEycTacWbouLChQu4uLjE65qyZctSrFgxVq5cmcTphJrCwsKoUr0KZXuUpWTzknFFWUI4aB2oZ1cPR3PHJEoZPwaDATc3NzJlysShQ4c+a1IhJiaGrFmzMnToUEaMGGHClO93M+Ym3q+8iVKiEvT91/z7DqeidUXKWpdFo9Hg6+uLh4fHW+eZmZnh5OSEn58fdnZ2Js2eVi1YsIBevXrx9OlTHBwc1I4jkoiseU6FjM+fE7ZoEYZ79xI9Ruz167xasgSjifY5FWlP69atyZ07d4K289LpdOzZs0dadadx1nbWDN45mBJNSwAfn+38kJfGl2wI3cC92MT/HjMFMzMzpk2bxpEjR/jzzz8/aywLCws8PT2TZe3/5ejLbA/bTqQSmeDvv/LvP8cjj3Mg/ACKotC5c+e44xqNBnNzcwwGA0FBQbKGNwH0ej0Gg0FadadxUjynMkpkJGHLlqGEh8PnFCiKgvHZM16tWIEidwaL97C0tKRv376sWLGCBw8exOsavV7Po0ePOHfuXBKnE2pRFIVdr3bx1OIpGm3iZ2kVFAwY2Bq6leeG5yZMmHBfffUVnp6eJmnb/abbYkIaDSXU7Zjb7H61O1FvWv5XYHQgR14eifuZzZ8/P40bN+aHH35g/fr1XL58mRo1anz246QXBQsWpFChQinixlGRdKR4TmUivL1RXr78vML5DUXB+PAhUamgK5hQR/fu3bG0tGTOnDnxOr9SpUrY2trKC0cadiX6CtdjrpukcFNQiCWW3a92q7ojh0ajYdq0aVy8eJE//vjjs8bS6XTExsYmWbfFKCUK71feJh3Tz+jHmVtnePXqFTdv3mTz5s2MGzcOLy8vaYKSCKlpz3OROFI8pyKxN24Q4+//4ZsCEynq2DEMDx+adEyRNmTKlImuXbvy22+/xauVsZWVFdWrV5cXjjQq0hjJgXDTfhytoPDA8IDAqECTjptQZcqUoXXr1owZM+az2nYXLlyYAgUKJNnPwPHw40QoESZ58/KGBg3nHc5jY2NjsjHTM71ez7Vr17hx44baUUQSkeI5FYk6fvz1rhmmptEQ5eNj+nFFmtCvXz9evnzJ0qVL43W+Xq/nyJEjREREJHEykdwuRl8kmuhPn5gIZ6LOqL4n/cSJE3n8+DGzZs1K9BhJ2W0x0hjJhegLJi2c4fUbmOfG59yOfX/7dZEw0qo77ZPdNlIJY0gIq7p3Z8aRIxiMRqJiY8lpb8+2b75BqzXBeyAzMzIOGoRG2q2K92jTpg0nT54kKCjok/vXXrx4keLFi+Pt7Y1er0+mhCIhYmNjmThxImvWrMHc3Bxzc3PKlStHvXr1GDt2LP7+/u9coygKf7z8g5fGlybJsHPKTiJeRNB0ctO4rzXO0Jj8FvnfOu/gwYNERkZSp04dkzzup3z//fcsXryY69evkz179kSNsWnTJpo3b86tW7dM2lQkd/7cxFrGYmFtgSHaQJUuVdCaaYkKjcKz7+ftw6xBQwHzAjS0b2iitOlb5cqVyZUrFxs3blQ7ikgCMvOcSvzj40P/HTtY0bIlR7/7jtN9+jBBrzfdXs0GA7GppK2uSH4DBw7kxo0bbN68+ZPnfvnll+TOnVuWbqRgnTt3xtfXlxMnTnD+/Hn8/PzQ6XQ8e/bsg9eEGkNNVji/jxYtt2Penfk8ePAgu3btSrLH/V8jRoxAo9Hw448/JnqMpGrVHUssHRZ3YMjhIXTf0J2/JvxFwXIFP7twhtezz7djb6fIbpCpkV6vZ9++fXF7Zou0Rf0d6kW83L9yBTOtlsz/WZPm+m8XN7+7dxm6cydh0dFYm5szqU4dKuTLR6zBQIvVq3kWHk5kbCwlcuRgdsOG2FlacuTGDYb8/TelcuXi3P37WJqbsyBPHsoXKaLWUxQpWJkyZahZsybTp0+nefPmH33TlpQfW4vPd+3aNTZs2MDt27fJnDkz8Pq/mZeXFwcPHiQ2NpaePXty7NgxYmNjWbZsGWXLluWh4SGn151m/6/7Acj8RWZa/NyCTLkzAbB39l7OrD+DRqvBwtqCXtt6YWlryf5f9uOzxgeNVkPu4rlp/lNzbDK+vbb23sV7bBi4ASLBOtqaNm3aMHLkSPz9/Zk/fz4Gg4GDBw/StGlTRo8ejbe3NxMmTCAiIgIzMzOmTp1qskYe2bJlY/jw4YwePZq+fftSqFChBI+ROXNmPDw82L1791tbwH0ORVEwKP9fiGXJmwXHIo4s77qcotWL0nRyU9YPWM9N35sAxETE8Pj6Y2Y9m8XvbX7n+Z3XO5pEvIhAo9Ew2n80Zzae4dD8QxhiDChGhXoj6hHSIoQsZllMkjk90+v1jB07Fl9f389uviNSHimeUwmXDBmokC8fJX/+mcoFClA+b16alyxJNltb2q9bx+yGDalVuDAnbt3im3XrONu3L3aWlixq1owstrYoisLAv/5i4alTfF+1KgCXHj9mSt26LGjalC0XLvDNmDFc7tRJOg+K9xo0aBD169fn6NGjVP3379CH6PV6/vjjD+7fv0+uXLmSKaGIj7Nnz1KkSBGyZcv23uOXL19m8eLF/Pbbb8yfP58RI0bg7e2N7zlfto/ZzsD9A8mUOxO7Z+xmbb+19NjQA581PpzbcY6+O/tik9GG8JBwzK3MubjnIidXnaT/7v7YOtiyrv86dozbQYsZLd56zCx5s9Bray/sre1pb9WeSpUqUbt2bSpUqECPHj0ICQmJW4ccHBzM2LFj8fb2JmPGjFy7do2qVaty8+ZNrKysTPI96tevH3PmzGHEiBGsXbs2UWPo9Xrmzp2LwWAwSavuaKLfWut87+I9Hl59SIk6JeK+1mLm6++rIdbA761/p2yLsgB0Xd0VgKiwKH79+leqda8GQDHPYrg3c0ej0fD09lNm6WfxfYPvyWIvxfPn8vDwwMHBgd27d0vxnAbJso1UQmswsKJlS3Z37kztwoU5efs2FebOJejpUzQaDbUKFwagYv78ZM+QgcAHD1AUhd9OnKDq/PlUnjeP3VevEvif/XrzZcpEdScnAJoUL87D58/5R5ZuiA+oW7cuLi4u/PTTT588t1atWgDs3bs3qWMJEytcuHDci33FihW5fv06AKcPn+bLWl/GzTRX6VyFoCNBGA1GLnhfoFLHSnEzyraZbNGaabl66CpuTdywdXjdybRyp8pcPXj1nceMiYxhbb+1jKs0jgoVKnDr1q33rrsG2LVrF9euXaNatWq4urrSvHlztFott2+b7mY3Gxsbxo8fz7p16xLdIOTNMhg/Pz+TZIpVXu8/vazzMqZVm8b679fT+tfWZHN6903QhoEbyJQrE18N/irua4ZYA0u/XUqp+qUo1+p16/Cnt56ywGsBUypNYXG7xYQ/D+fWjVsmyZvemZubJ1vDHJH8pHhOLcxff0hQNHt2vi1bltWtW+ORJw87Ll1659Q388YbAgM5fOMGf3XsyPGePeldqRKRH2oAoNGg+fd/QryPRqNh0KBBbN++nStXrnz0XEdHR9zc3OSFIwVyd3cnKCiIp0+fvve49X9uGjYzM4trGqL935eLxPyq+MA1f034C7ssdow8PJKAgABq1KhBZGTke89VFAWdToe/v3/c/+7evUsREy8569ChAyVKlGDIkCGJ2gWkQoUKZMiQwWQ/A+b/flD8Zs1zf+/+uDZyfec87+nehNwLwWum11tf3zBgA5lyZ0I/6P9v4l3eZTkV2ldg2PFhDDk8BCs7K2KiYkySV7z+9OHEiRO8fJl09woIdUjxnEo80Gg4+Z9Z4ZCICG6FhFAyZ04UReHAv7NDp27f5lFYGCVz5iQkMpIstrZktLYmNCqK1f8zk3M7JITD/+5Due3CBRyzZCFPnjzJ9pxE6tOmTRty5swZr5bder2ePXv2qL79mHhb4cKFadasGZ07dyYkJAR4XZBu2rSJ4ODgD15Xs2ZNLu27xIv7LwA4vvQ4RasVRWumpUTdEhz/4zgRL19vTxj+IhyjwUjR6kXx3+pP5MvXhfDxP47jXPPdphvhIeFkyp2JbJbZuHLlyls32mXMmJEXL17E/fmrr75i7969b3Wx9EmCrTbNzMyYMmUKBw8eZOfOnQm+3tStui01lmg+8Y7FZ40PgX8H8u3SbzEz//+lIrum7eLFgxd4zXi7oA4PCSdr/qwA+K73JTwknIzajCbJK15/+vBmvb5IW2TNcyqhZMvGtIMHufX8OTYWFhiMRlqXLk39YsXIbW/P0J07GeHtjbW5OctatCCDlRWtSpfm78uXKfvrr2SztaVi/vz88++LJcCX2bOz2t+fYTt3YmFmxop58zAajfzzzz9cu3aNoKCguP9v164dLVq0+HBAkS5YWVnRt29fxo0bx4QJE3B0dPzguTqdjqlTpxIYGEipUqWSMaX4lCVLlvDjjz9Svnx5zM3NMRqNVKtWjbp1637wmiqlqtBwXEPme80HXt8w2HJWSwA8Wnrw4sELZn81G625FktbS3pu6YmLzoUHlx4w66tZb90w+L/0g/Ss6rGKwHWBlCpSCk/P/989okmTJqxYsQJXV9e4GwZXr15N9+7dCQ8PJzo6Gjc3N1avXm3i7xLUq1eP6tWrM3ToUL766qsEr13W6XQMGDCAsLAwMmTI8FlZNBoNZpqPP/6uqa93JZldb3bc14YcHsKuKbvIXjg7MzxnAOCQ04Hu67vTdHJTlnZcio2DDUWqFiFznsxSPJtQoUKFcHJyYvfu3TRsKFsApiWyz3MqYXz+nNBffjHZeEdu3GD4rl0c/e67118wM6Puli2c9POLmynUaDRxH9sOHz6cSZMmmezxRer17Nkz8uXLx4ABAxg/fvwHz4uMjCRLliyMHz+eQYMGJWNCkRQURWHpi6WEKqFJ9hiNMjSigEWBJBs/MXx8fChfvjxLlizh22+/TdC1V69exdnZmT///JP69et/dpazkWc5EnHks8d5Hw0a8pnno7F94yQZP7367rvv2LdvH1evvrvWX6ResmwjldBmzoyZk1PSdBjUarEoWRIXN7e3PmJXFCVuvWPz5u/OFon0KUuWLHTu3Jm5c+cSHh7+wfOsra2pXr26dNlKIzQaDaWtSyfZ+Bk0GchnbrqGIqZSrlw5WrRowahRoxLcNbNIkSLkz5/fZEs3vrT88t215yaioCTpf9/0SqfTERQUxM2bN9WOIkxIiudUxKpSJTDRBwVVCxb8/1lnoxGr8uVZtGjRB2eXx44dy6pVq+TGBwFA//79CQkJ4Y8//vjoeXq9nsOHD0ur7jTCxdIFCyySZGx3a3e0mpT5kjRp0iQePXrE7NmzP33yf2g0GnQ6ncneQNpobXCxdPnk2ueE0qDBQetAfvP8nz5ZJEhSNcwR6kqZv6nEe1kUKoRFiRKmnX3WaLCsWBGznDkB3lmeYWZmhqenJ48fP6Zdu3Y4OjrSuHFjKaTTuYIFC+Ll5cXMmTM/2kFLp9MRGRnJ0aNHkzGdSCo2Whuq21Y36ZgaNGTXZidHSI4Ue3NpoUKF6NGjB1OmTPngLiUfotfruXTpksm2Aa1sUxlrjfWnT0wABQW9nT7FvnlJzTJlykT58uVl56E0Rn5SUhnrevXQZMhgmgJaq0WbNSvW/9OZa/jw4UycOBEAg8HA9OnTOXHiBLdu3WLSpEk8fPjwrUJ69erVhIYm3TpIkTINGjSI69evs23btg+eU7x4cXLlyiUvHGmIi6ULBc0LmmT2U4MGM8zYN24fX+T6Ant7e6pVq0bv3r2ZP38+R48eTTG/W0aNGoXRaIz73RhftWrVQqPRmGzm0Vprjd5O/+kTE6CsdVlym+c26Zji/+l0OmnVncbIDYOpkOHpU14tXYoSEQFGY+IG0WrRODiQ4dtv0drbv/eUGTNmcPjwYbZu3frO/s+3b99m48aNrF+/nlOnTmFlZUXdunXx8vLi66+/xv4DY4q0pUaNGkRHR3P8+PEPntOxY8e4/XhF2hCjxLDuyTqeaJ+g0SauiNagQYuWxhkac8/v3ltd2CwsLIiNjUVRFIoXL8758+dNFf2zTJw4kXHjxnHlyhUKFiwY7+vKly+Pk5MTa9asMVmWi1EX2RP++QW5i6ULtW1ryx7/SejYsWNUqVKFU6dOUa5cObXjCBOQmedUyCxrVjJ07oxZjhyJHyN/fjJ07vzBwhlg4MCBbNu27b2/VN/stnDy5Elu3rzJxIkTuX//Pm3btiV79uw0adKENWvWpJhZI5E0Bg0axIkTJz5aPOt0OgICAnjwn+6WInWLiYhh9tezubDjQqKu16AhgzYDze2bk8ciD+XKlaNt27Zota9fkmJiYuKWcLRq1cpkuT9X//79yZYtGyNHjkzQdW/WPRsTO9nxHi5WLjSwaxCv/Z//l+bff8pbl5fCORmUK1eOjBkzyidwaYjMPKdiitFI9IkTRB46BDGf6Aql0by+2dDaGpvatbFwd0+SX5i3bt2Km5H28fHB2to6bka6QYMGMiOdxhiNRooXL06xYsXYsmXLe895+PAhOXPmZMWKFbRr1y6ZEwpTUxSF1q1bs2PHDo4fP05Gl4zsf7WfMCUMDRoUPvyS8qZoK2VViko2lbDQ/P/Nh9euXcPZ2fmtAtPZ2Znz589jbp5yWhL8/vvvdOvWDV9fX8qUKROvaw4fPkz16tUTdE18vTK+4lD4IYJiguL1/VdQyGaWjdq2tclhnvgJGJEwjRs35vnz5xw6dEjtKMIEpHhOA5ToaGICA4kOCMBw/z78bwtuS0vMcufG0s0NCxcXNMn0QnTz5k02btzIhg0b3iqkW7RoQYMGDT67aYBIGRYtWkS3bt24fPkyRYsWfe85bm5ulCpVimXLliVzOmFqkyZNYsSIEWzcuJFmzZoBYFSM3Iq9xfmo89yLvUek8nZrbS1asphlwdnSGRdLF2y1tu8du3Pnzixbtgyj0YiFhQXR0dG0adOG+fPnp5g33rGxsZQqVYrcuXOzZ8+eeE1CREdHkzVrVkaMGMGwYcOSJNcLwwsCowIJjgnmufH5O8ftNHbkNc9LSeuS5DLLJbPNyey3336jX79+PHv2LMX8XRaJJ8VzGqMYjRifPXu9HlqjQWNrizZzZtV/Ub6vkK5Xr17cjLQU0qlXZGQkBQoUoEmTJsybN++95wwZMoQVK1Zw79491f8uisTbsWMHjRo1YtSoUYwbN+695yiKwivlFWHGMIwYsdRYklmb+ZPd8eD1J1eFCxcmNjaWv//+m5CQELp3706OHDlYu3atyWdtE2v79u00atSIXbt28dVXX8Xrmq+//ppXr16xf//+JE73ek36c8NzYolFixYHrQM2Wpskf1zxYdeuXaNIkSJs376dr7/+Wu044jNJ8SyS3ZtCev369Zw+fVoK6TRg4sSJ/Pjjj9y+fZvs2bO/c3zv3r3odDrOnTtHyZIlVUgoPtfFixepUKECtWvXZuPGjXHrk03tzV7K/fr1A14XHa1ateLcuXNMmzaNfv36qf4GTFEUqlWrxsuXLzl79my82nb/+uuvDBw4kOfPn2NnZ5cMKUVKoigKTk5OfP311/xiwm7BQiWKECoKDg5Wpk2bpnh4eCiAYm1trTRt2lRZu3atEhoaqnY8EU9PnjxRbG1tlbFjx773eEREhGJtba3MmDEjmZMJU3j69KlSqFAhpWTJkqr8XEZFRSkDBgxQAKVBgwbK48ePkz3D/zpx4oQCKMuWLYvX+ZcvX1YA5a+//kriZCKl6tatm+Ls7Kx2DGECMvMsUowbN27EzUj7+vpiY2MTNyNdv359mZFO4fr06cPatWu5ffs2NjbvfkT81VdfodFo2LVrlwrpRGLFxsZSt25d/Pz8OH36dIK2aDO1v/76iw4dOmBlZcXq1aupXt20DVsSqnnz5vj4+HD16lWsrT/euERRFPLnz0+zZs34+eefkymhSEk2bdpE8+bNuXXrFvnypbxW9CL+ZKs6kWIULFiQwYMHc/r0aYKDgxk7diy3bt2iVatWODo64uXlxfr163n16pXaUcV79O/fn2fPnrF8+fL3Htfr9Rw6dIjIyMj3Hhcp0+DBgzlw4AAbNmxQtXAGqF+/PgEBARQtWhRPT0/Gjh2rauOJSZMmcf/+fX799ddPnqvRaNDr9bJdWTomrbrTEJVnvoX4pOvXrytTp05VypQpowCKjY2N0rx5c2XdunVKWFiY2vHEfzRv3lwpUqSIEhsb+86xc+fOKYCyd+9eFZKJxFiyZIkCKHPmzFE7yltiY2OV8ePHK1qtVqlWrZryzz//qJalZ8+eSqZMmZSnT59+8tx169YpgKp5hbrKly+vtGjRQu0Y4jPJzLNI8ZycnBgyZAi+vr5cv36dMWPGcOPGDVq2bEn27Nnx8vJiw4YNMiOdAgwaNIigoCB27NjxzrESJUqQM2dOmXlLJU6cOEGPHj3o2rUrPXv2VDvOW8zMzBg1ahQHDhwgODiY0qVLv/fvXHIYPXo0sbGxTJ48+ZPnvmnVvXfv3mRIJlIivV7P3r17pVV3KidrnkWqdf369bg10mfPnsXW1pb69evj5eVFvXr15I52lVSrVg2j0cjRo0ffOfbNN98QGBiIn5+fCslEfN25c4eyZctSpEgR9u3bh6WlpdqRPujp06d06tSJ7du3069fP6ZOnYqVlVWyZhg/fjwTJ07k6tWr5M+f/6Pnenh4UKRIEVavXp1M6URKcvToUapWrcrp06cpW7as2nFEIsnMs0i1ChUqxNChQzlz5gzXrl1j1KhRXL9+nRYtWuDo6EiLFi3YuHEj4eHhakdNVwYNGsSxY8c4ceLEO8f0ej3+/v48fPhQhWQiPiIiImjcuDGWlpZs2rQpRRfOAFmzZmXr1q3Mnj2befPmUalSJYKCgpI1w4ABA8icOTOjRo365Ll6vd7krbpF6lG+fHns7e3lE7hUTopnkSYUKlSIYcOGvVVIX7t2DS8vL7Jnz07Lli2lkE4mDRo0oGjRosyYMeOdY7Vr1wZg3759yR1LxIOiKHTp0oWLFy+ybds2HB0d1Y4ULxqNhr59+3LixAlCQ0Nxd3dn5cqVyfb4GTJkYOzYsaxcuRJ/f/+PnqvX63ny5AkBAQHJE06kKBYWFtSsWVOK51ROimeR5rwppM+ePUtQUBAjR44kKCjorUJ606ZNUkgnEa1Wy8CBA9m8eTPXr19/61jOnDkpVaqUvHCkUNOnT2f16tUsXboUNzc3teMkmLu7O2fOnKFJkya0b9+ejh07EhYWliyP3blzZ4oWLcrQoUM/el7FihWxs7OTn4F0TK/Xc/z48WT7uylMT4pnkaYVLlyY4cOHv1NIN2/eHEdHR1q1aiWFdBJo37492bJle+9+tm+265LbLVKWv//+m2HDhvHDDz/QsmVLteMkmr29PcuXL2fZsmVs3LiRsmXLfnI22BQsLCyYPHkyu3fv/uhWZJaWltSoUUOK53RMp9MRExPDoUOH1I4iEkmKZ5Fu/G8hPWLECK5cufJWIb1582YiIiLUjprq2djY0Lt3b5YsWcKTJ0/eOqbX67l//z4XL15UKZ34X5cvX6Z169Y0aNCACRMmqB3HJL755hvOnDmDjY0NFSpUYO7cuUn+hq1x48ZUqlSJoUOHfnRNs16v5+jRo/KmPZ0qUqQI+fPnlzdQqZgUzyJdelNI+/n5cfXqVX744QeuXLlCs2bNyJ49O61bt5ZC+jO92d5s3rx5b329SpUqWFlZyQtHChESEkKjRo3IkycPK1euRKtNOy8Lzs7OnDhxgm7dutG7d2+aNm3Ks2fPkuzxNBoN06ZNw8/PjzVr1nzwPJ1OR3R0NIcPH06yLCLlkoY5qZ9sVSfEfwQFBbFhwwbWr19PQEAAdnZ2fP3113h5eVG3bt33tp0WH9arVy82bNjA7du332pfrNfrMTMzY+fOnSqmEwaDgQYNGnDq1Cl8fHwoXLiw2pGSzLZt2/j222/JkCEDa9asoXLlykn2WE2aNMHPz48rV668d9s8RVHIly8fXl5ezJw5M8lyiJRrw4YNtGjRgtu3b5M3b16144gESjtTDEKYQJEiRfjhhx/w9/fnypUrDB8+nEuXLtGsWTMcHR1p06YNW7ZskRnpePr+++958uQJK1aseOvrb1p1R0VFqZRMAAwbNozdu3ezbt26NF04AzRq1IiAgADy589P9erVmThxYpI1qpg8eTJ37txh7ty57z0uM4/iTcMcadWdSqnX3FCI1OPKlSvKjz/+qJQqVUoBlAwZMiitW7dWNm/erISHh6sdL0Vr2rSp4uzsrBgMhriv+fv7K4Cyb98+FZOlb8uXL1cAZdasWWpHSVYxMTHKqFGjFI1Go3h6eir37t1Lksfp3r27kiVLFuX58+fvPb5mzRoFUO7evZskjy9SvnLlyiktW7ZUO4ZIBJl5FiIeihYtyogRIwgICODy5csMHTqU8+fP07RpUxwdHWnbti1bt24lMjJS7agpzqBBg7hy5Qp//fVX3NdKlixJjhw5ZOZNJT4+PnTt2pVvv/2Wvn37qh0nWZmbmzN+/Hj27dvHpUuXKF26NLt27TL544wZM4bIyEimTJny3uO1a9eWmcd0TqfTsXfvXmmYkwrJmmchPsOVK1fi1kgHBgZib28ft0a6Tp06b63zTc8qV66Mubn5W1sztW/fnosXL3LmzBkVk6U/9+7dw8PDg/z583PgwIFkb2Wdkjx+/JgOHTqwc+dOBg0axMSJE03aUXHMmDFMnTqVoKCg965rLVu2LM7OzqxatcpkjylSj8OHD1O9enV8fX0pU6aM2nFEAsjMsxCfwdnZmZEjR3Lu3DkuXbrE4MGDCQwMpEmTJnEz0tu2bUv3M9KDBg3i8OHD+Pj4xH1Np9Nx9uxZHj9+rGKy9CUyMpImTZqg0WjYvHlzui6cAbJnz86ff/7JjBkzmD17NlWrViU4ONhk4w8aNAgHBwdGjx793uMy85i+VahQgQwZMsgncKmQFM9CmEixYsUYNWrUW4X0uXPnaNy4MY6OjrRr1y7dFtINGzakSJEi/PTTT3Ff0+l0AOzdu1etWOmKoih0796dc+fOsXXrVnLmzKl2pBRBq9UyYMAAjh07xpMnT3Bzc2PdunUmGdve3p7Ro0ezbNkyzp07985xvV7Po0eP3ntMpH1vGubI0p3UR4pnIZLAm0I6MDCQixcvMmjQIAICAt4qpLdv355uCmkzMzMGDBjApk2b4mb2cuXKRcmSJeWFI5n8/PPPLF++nMWLF1O2bFm146Q4Hh4enD17lrp169KqVSu6detmkiYm3bp1o3DhwgwbNuydY5UqVcLW1lZmHtOxNw1zXr16pXYUkQBSPAuRxL788ktGjx79ViHt7+9Po0aNcHR0pH379umikP7mm2/IkiULs2bNivuatOpOHt7e3gwePJghQ4bQpk0bteOkWA4ODqxZs4ZFixaxcuVKPDw8OH/+/GeNaWFhwaRJk9i5cyf79+9/65iVlRXVq1eXN5DpmF6vl1bdqZAUz0IkozeF9Pnz57lw4QIDBw7Ez8+PRo0akSNHDtq3b8+OHTvS5P7Htra29OrVi8WLF8d1edPpdNy9e5dLly6pnC7tCgoKolWrVtSpU4dJkyapHSfF02g0dO7cGV9fX7RaLR4eHixcuPCz3uA1a9aM8uXLM2TIkHfWN+v1eo4cOSKtutOpokWLkjdvXnkDlcpI8SyESlxcXBgzZkxcIT1gwADOnj1Lw4YNcXR05JtvvklzhXTPnj0xGo3Mnz8fgKpVq0qr7iT04sULGjZsSI4cOVi9ejVmZmZqR0o1XFxc8PHxoWPHjnTv3p2WLVsSEhKSqLHetO0+c+YM69evf+uYXq8nKiqKI0eOmCC1SG2kYU7qJFvVCZHCXLhwIW77u0uXLpExY0YaNWpEixYt0Ol0qX6HhB49erB161Zu3bqFlZUVOp0OS0vLt/aBFp/PYDDQqFEjjh49io+PD0WLFlU7Uqq1ceNGunTpQubMmVm7di3ly5dP1DgNGzbk/PnzXLp0Ke7nWFEU8ubNS6tWrd66oVakH+vXr6dly5b8888/5MmTR+04Ih5k5lmIFKZ48eKMHTuWixcvcv78eb7//nt8fX35+uuvyZEjBx06dODPP/9MtTPSAwYM4NGjR6xcuRJ4vXTj4MGDqfb5pFQjR45k586drF27Vgrnz9S8eXP8/f3JmTMnVapUYdq0aYnaXm7KlCncunUr7pMXeD3zqNPpZOYxHXvTqlt2HkpF1GtuKIRIiPPnzyujR49WihUrpgCKg4OD8s033yg7duxQIiMj1Y6XII0bN1a+/PJLxWAwKH5+fgqg7N+/X+1Yacbq1asVQJk+fbraUdKU6OhoZdiwYQqgfPXVV8qDBw8SPEaXLl2UrFmzKiEhIXFfe/PfK6lahYuUr2zZskrr1q3VjiHiSWaehUglihcvzrhx47h48SKBgYH069cPHx+fuBnpjh078tdffxEdHa121E8aNGgQly5dYufOnZQqVQpHR0e5YcZEzpw5Q6dOnWjXrh0DBw5UO06aYmFhweTJk/H29sbPzw9XV9cEzxaOGzeO8PBwpk2bFve12rVrA8jPQDqm1+vZs2ePNMxJJWTNsxCpmKIoXLhwgfXr17N+/XquXLlCpkyZaNSoEV5eXnHriVMaRVGoVKkS1tbWHDhwgLZt23LlyhV8fX3VjpaqPXjwAA8PD3LlysWhQ4ewsbFRO1Ka9eDBA9q3b8++ffsYPnw4Y8eOxcLCIl7Xjhw5kpkzZxIUFMQXX3wBgLu7O8WLF2fFihVJGVukUIcOHaJGjRqcPXsWNzc3teOIT5CZZyFSMY1GQ4kSJRg/fjyXLl3i3Llz9OnTh5MnT9KgQYO4Gem///47Rc1IazQaBg0axMGDB/H19UWv13P27FmePHmidrRUKyoqimbNmmEwGNiyZYsUzkksZ86ceHt7M2nSJKZOnUqNGjW4detWvK4dMmQIdnZ2jBkzJu5rMvOYvlWsWBE7OztZ+55KSPEsRBqh0WgoWbLkW4V07969OXHiBPXr1ydHjhx8++23KaaQbty4MU5OTsyYMQOdToeiKOzbt0/tWKmSoij06tWLM2fOsGXLlrjZTJG0tFotw4YN48iRI9y9exdXV1c2b978yesyZszIqFGjWLp0KRcuXABeF88PHz4kMDAwqWOLFOhNq24pnlMHKZ6FSIPeFNITJkzg8uXLBAQE0Lt3b44fP/5WIb1z507VCuk3Lbs3bNhAdHQ0xYsXlxeORJozZw6LFy9m4cKFid5GTSRexYoV8fPzo1atWjRr1oxevXp9smNojx49KFCgQFzb7sqVK2NjYyPrntOxN626pWFOyidrnoVIRxRFITAwMG6NdFBQEJkyZaJJkyZ4eXlRq1atZF0jHR4eTr58+WjXrh1arZYNGza8/uj75UuU6GjQatFmzIgmBa7bTin27dvHV199Rb9+/ZgxY4bacdI1RVFYsGAB/fv3x9nZmbVr1/Lll19+8Px169bRqlUrDh48SPXq1albty4Gg4Hdu72BWEABNID5v/8v0rJLly7h4uLCzp07qVOnjtpxxEdI8SxEOqUoCufOnYtryBIUFETmzJlp3LgxLVq0oFatWvG+AepzjB49msVz57J54kSe+fhQ0ckJjcHw1jnaLFkwK1AAS3d3zGVJQpzr16/j4eGBh4cHf/31F+bm5mpHEsC5c+do1aoVt27dYs6cOXTs2BGN5t3i12g0Ur58ebRaLSdPHuXkyX1YWsbi7l6Sd0+3AGyADP/+u0hrFEUhX758eHl5MXPmTLXjiI+Q4lkIEVdIv5mRvnbtGpkzZ35rRjopCmklJoZnf/2F4cwZLP8t/D44v6bVgtGIWa5c2DRogFnu3CbPk5qEhoZSsWJFoqOjOXXqFJkzZ1Y7kviPV69e0a9fPxYvXkybNm2YP38+9vb275x39Oghbt70o23bxoDmPUXz+1gDWXk9Iy3Sks6dO3Pq1CnOnz+vdhTxEVI8CyHeoigKAQEBcTPS165dI0uWLHEz0p6eniYppGPv3iV840aUFy8gIb+G/q0urKpWxapGjffO6KV1RqORpk2bsn//fk6dOvXRpQFCXWvWrKF79+7kyJGDtWvXUqZMmf8cDQeeYjTGotUm9BYkDZAZeLcgF6nX2rVrad26NXfv3iV3Op8gSMnkhkEhxFs0Gg2urq5MnDiRq1ev4ufnR/fu3Tl8+DB16tQhZ86cdOnSBW9vb2JiYhL1GLE3bvDqjz8SXjjD6/MVhajDh4nYvBklHW7tNWbMGLZv386aNWukcE7hWrdujZ+fH5kyZaJixYrMmjWL13NWYcBjwJiIwhler4d+BoSYMK1QW+3atdFoNHLjaAonM89CiHhRFAV/f/+4Genr16+TJUuWuKUd8Z2RNjx8SNiiRWAwJLxwfg/LcuWwqVv3s8dJLTZs2ECLFi2YPHly3E4NIuWLjo5m+PDhzJw5k3HjhjFqVI94LtGIjyzIDHTaUbZsWZydnVm1apXaUcQHSPEshEiwN4X0+vXr2bBhQ1wh3bRpU7y8vKhZs+Z7C2nFYCBs4UKMjx+bpHB+w7ZdOywKFTLZeCmVv78/lStXplGjRqxatSpdLllJ7fbs2UXZsrnJlCmjif/75UZuJEwbhg8fzpIlS7h//34iP5UQSU2KZyHEZ1EUBT8/v7gZ6eDgYLJmzRo3I/3fQjry6FGiTN0IRaNBkyED9n37oknDu008evQIDw8PsmXLxpEjR7C1tVU7kkiUxyhKuAlnnd+wAnKaelChggMHDuDp6Ymfnx+urq5qxxHvIcWzEMJk3hTSb2ak/1tIt2zeHI+AAIiISJLHtmnSBMtSpZJkbLVFR0dTu3Ztrl69yunTp8mbN6/akUSixAJ3k3D8nLwuokVqFhUVRZYsWRgzZgxDhgxRO454D/k8QAhhEgUKFCAgIAB3d3emTJnCtWvX8PX1pUuXLuzbt485AwYkqHD+Zt06Vvn5ATBx/37WnzsHwOQDBxi2c+fbJ2s0RJ86xdixY+nfv7+pnlKK0bdvX06ePMmmTZukcE5lNm/eTJkyZXB1daVYMRc8PVtjTLKbXEMBGDt27Cc7HIqUy8rKiurVq8tNgylY2v2MUwihKo1GQ5kyZShTpgyTJ0/mn6VLMf7zT6LesY/w9Pz4CYqC4d49lNjYRGVNyebNm8eCBQtYtGgRlStXVjuOSID79+/TrVs3zpw5Q/78+YH7nD17NgnXqr9+czpu3Dj69++PtbV1Ej2OSGp6vZ5hw4YRHh4uS7RSIJl5FkKYVI0aNRg0aBBVq1alUKFC9OjRA41GQ+bISF5FRdF3+3Y8Fy6k0m+/0W/7dqL/LXivPn6MftEiKsydS5s1awiNioob87stW/jtxIm4P999+ZKv//gDj19/peXq1TwLDwdACQt7K8tPP/1EuXLlcHd3p06dOq9bf6cihw4dom/fvvTt25fOnTurHUck0MOHDzEzMyNLliy83louGnf3EuzZcwS9vj0AL1+GYmFRmIULVwOwfPkmOnUaDMCDB49o0aIX5co1omTJrxg58qe4sYOCblC//rd4eDSkVKk6zJmzDDDSo0d3AKpWrYqrqyuPHj1KzqcsTESv1xMVFcWRI0fUjiLeQ4pnIYTJXb9+nQMHDnD+/Hm8vb05ceIExmfPGOntTcV8+djfrRvHvvsOo6Iw/9QpALpv2UI7d3dO9urFCE9Pjn2k0D1x6xaLmjXjdJ8+5MmYkXF79wJg/LeIBli9ejVXrlzhxIkTnD17lrZt29KzZ8+kfeImdPPmTZo3b0716tWZMWOG2nFEIpQqVYoqVaqQP39+mjRpwvTpC7h79wFVq5bD1/ccUVFRHDhwAg+PUuzdewyAPXuOUrt2FQA6dBhIr17t8fHZhp/fX/j6nmPDhr8wGAy0bt2XGTNGcPr0dk6e3MLChWs4fTqA+fNnA3DkyBH8/f1xdHRU7fmLxPvyyy/JnTu3LN1IoWTZhhDC5Fq2bIm5uTnm5ua4urpy7do1XAwG/rp8mdN37jD331nkyNhYzLRaXkZGEvjgAW3/vbO8eI4cVMiX74Pj64sWJce/rY47lClD+3XrXnce/M9a0q1bt3L69Om4jm4GgyGJnq3phYWF0bBhQzJmzMi6deswT8O7iKRlWq2WTZs2cfnyZQ4d2s/OnVuZOHEOvr47cHV14dixM+zde4xhw75jwIAfMRqN7N9/nOnTh/PqVTj79h3n4cMnceOFhYVz5UowV64Ec+HCVVq16hN3LDQ0jIsXg/Dw0KnxVIWJaTQa9Ho93t7e/PTTT5++QCQr+Y0shDC5/661NDMze124arUowPIWLSicLdtb5798z81NCVkVqtFoXu8b/Z+1pIqiMHz4cLp165bQ+KoyGo106NCBGzducPLkSbJmzap2JPGZihUrRrFihejevT516nzD9u17qV27Cnv3HuXw4VNMmTKUkiWdWblyC5kzO5AzpyNhYa8AOHlyyztrly9cuEqWLJnw99/5nkeTvb/TCr1ezx9//MH9+/fJlSuX2nHEf8iyDSFEstBmzkz9YsWYdewYsf/OAodERBD89CkZra0plTMnawICALj06BEnb9/+4Fh7goJ49O/65hVnz1LDyen1Y/znxprGjRszf/58nj17BkBMTAx+/+7ekZJNmDCBzZs3s3LlSooXL652HPEZ7t69y7Fjx/79kznPn7/gxo07FCqUj9q1K7N69TYyZcqInZ0ttWtXYfTomdSu/fqm0AwZ7KhZsyJTpsyLG+/evYfcuXMfZ2cnMmbMwNKl6+OOXbt2k2fPQgBz7O3tefHiRfI9UZEkateuDcDef5eliZRDZp6FEMnCLE8eJtWty7jdu6k6fz4ajQZzrZbxOh1OWbMyv0kTem3bxtzjx3HKmpVK+fN/cKyK+fLRZdMm7r98iVPWrMxr3BgATYYM8Or1jF3btm15+vQpNWvWBCA2NpZOnTrh5uaW5M81sTZv3szYsWOZMGECjRo1UjuO+EyxsbGMHz+eGzduYGtrS2xsJB06NKNRIz1Go5EXL0KpVet1sazTVaF379FxfwZYtWoWAwb8SIkSejQaDXZ2NixYMIk8eXLx559L6N9/PD//vASDwUC2bJlZvXoOYM7AgQPR6XTY2tqye/duWfecSmXPnh03Nzd2795N+/bt1Y4j/kOapAghkkV0QAARW7cm2fhaR0fsv/suycZPaoGBgVSsWJF69eqxbt06ab2dJj0HXibh+LZA9iQcXyS3YcOGxS3dkN8JKYcs2xBCJAsLFxewSrruZ1blyiXZ2EntyZMnNGzYkCJFirB06VJ5kUyzMiTx+PZJPL5Ibnq9nocPHxIYGKh2FPEfsmxDCGEyoaGhdOnShYcPHxIdHU10dDQxMTE8e/aMkJAQjk6bRoHHj1/f3GdCGmtrLEqWNOmYySUmJgYvLy9evXrFwYMHsbOzUzuSSDIWgA1vmpmYfmxpzZ3WVK5cGRsbG3bv3k2pUqXUjiP+Jcs2hBAm8/LlS3LmzEnEe9pwa7Vazp46RcFjx1BevDBpAW3r5fV6ZjsV6t27NwsWLGD//v1UrVpV7TgiycUC93jdNMV0wsMdsLFxkE8t0qC6detiMBjYvXu32lHEv2TZhhDCZDJmzMjgwYPf+wK+ZMkSSpcti23Tpm9tKfdZNBosSpRItYXz77//zty5c5kzZ44UzumGOZDFpCM+e2bEzi4zVlZWeHh40K1bN3755Rf279/P48ePTfpYIvnp9XoOHz783kkJoQ6ZeRZCmMzFixcZNmwYO3bsiPuamZkZzZs3Z82aNXFFdczFi4Rv3Ph5s88aDWYFC2LXujWaVNhE5OjRo3h6etKlSxd+++03teOIZPcCCDHBOBkwGjPxxRd5ePDgAQDm5uYYjUaMRiNmZmb4+/tTokQJEzyWUMP58+cpWbIku3fvRqeTJjgpgcw8CyE+240bN+jQoQMlS5YkMDCQli1botFo0Gq15MqViwULFrw1G23h4oJtmzZobGwSPgv97/kWbm6ptnC+ffs2TZs2pVKlSsyePVvtOEIVDnz+DHQmIAtarRkLFy6M+2psbCxGoxGtVkvRokUpVKjQZz6OUFPx4sXJlSuXLNtIQWTmWQiRaA8ePODHH39k4cKFZM2alVGjRtGlSxdiY2MpUKAAT58+5ejRo1SsWPG91xvDw4nctYuYwMDXRfHHfh39e1zj4IDN119jkUoLgvDwcKpUqcKzZ884ffo02bPL1mLpWwzwFIhKwDUWQDbAMu4riqJQtmxZAgIC4lrRm5mZ4ePjg7u7uwnzCjV07NgRPz8/Av5tJCXUJcWzECLBnj9/zrRp05g9ezbW1tYMHTqUPn36YPufDn9Hjhzh8ePHNG3a9JPjGUNCiPb1JebiRYzPn797gpUV5vnyYVm2LOaFC6PRps4PzRRFoXXr1uzYsYPjx49TunRptSOJFCMaCOX1ThyG9xzXAta83o7Oive14fb29qZOnTpxf7awsKBo0aKsW7dOulWmcqtWraJdu3bcv3+fnDlzqh0n3ZPiWQgRb2FhYfzyyy9MmzaN2NhYvv/+ewYOHEimTJlM9hhKVBSGx48hOhq0WrQODmgyZUoTuwhMmjSJESNGsHHjRpo1a6Z2HJFiGXg9I63wukg2Jz47yyqKQuXKlTlx4gQtW7Zk1KhRtGrVimvXrvHLL7/QpUuXNPFzlB49fPiQnDlzsmLFCtq1a6d2nHRPimchxCdFRUWxYMECJk6cSEhICN999x3Dhw8nR44cakdLNXbs2EGjRo0YPXo0Y8eOVTuOSKPOnDnD1KlTWbRoERkzZiQiIoIBAwYwf/58WrRowcKFC3FwcFA7pkgENzc3SpYsyfLly9WOku5J8SyE+KDY2FhWrFjB2LFjuXPnDh07dmT06NHkz59f7WipyoULF6hQoQI6nY6NGzeiTaXLTkTqtXHjRrp06ULmzJlZu3Yt5cuXVzuSSKChQ4eyfPly7t27J58gqEx+gwsh3mE0GtmwYQMlSpSgU6dOlC9fngsXLrB48WIpnBPo2bNnNGrUiIIFC7J8+XIpnIUqmjdvjr+/Pzlz5qRKlSpMnz4do9GodiyRADqdjgcPHnD+/Hm1o6R78ltcCBFHURR27dqFh4cHLVq0wMnJiTNnzrB+/XqKFSumdrxUJzY2lpYtWxISEsK2bdvIkCGD2pFEOlagQAEOHz7MoEGDGDJkCPXq1ePRo0dqxxLxVKVKFaytrWXLuhRAimchBADHjh2jevXq1K1bF1tbWw4fPszff/8t21x9hkGDBnHgwAE2bNhAwYIF1Y4jBBYWFkyePBlvb2/8/PwoXbo0e/fuVTuWiAdra2uqV6/Onj171I6S7knxLEQ65+/vT/369alSpQqhoaH8/fffHD58WNpFf6alS5cye/ZsZs+eTc2aNdWOI8Rb9Ho9AQEBlChRAr1ez4gRI4iNjVU7lvgEnU7HoUOHiIyMVDtKuibFsxDp1NWrV2nVqhVubm5cu3aNdevWcebMGerWrSs3o3ymEydO0KNHD7p27UrPnj3VjiPEe+XMmRNvb28mTZrE1KlTqV69Ordu3VI7lvgIvV5PZGQkR48eVTtKuibFsxDpzD///EPXrl1xcXHh+PHjLFq0iAsXLtCiRQu5mc0E7ty5Q5MmTShXrhxz5syRNyIiRdNqtQwbNowjR45w9+5dXF1d2bx5s9qxxAeUKFGCnDlzyrpnlckrpRDpxKNHj/j+++8pXLgw27ZtY8aMGVy9epXOnTtjbv7pBgzi0yIiImjcuDGWlpZs2rQJS0vLT18kRApQsWJF/Pz8qFWrFs2aNaNXr16yNCAF0mg06HQ6WfesMimehUjjXrx4wejRoylUqBBLlixh1KhRXL9+nX79+mFtba12vDRDURS6dOnCxYsX2bZtG46OjmpHEiJBMmfOzIYNG5g3bx6LFy+mfPnyXL58We1Y4n/o9Xr8/f15+PCh2lHSLSmehUijwsPDmT59Ok5OTkyfPp2ePXty48YNRo4cib29vdrx0pzp06ezevVqli5dipubm9pxhEgUjUZDjx498PHxISYmhjJlyrB06VKkn1rKUbt2bQDZJUVFUjwLkcZER0czb948ChcuzA8//EDLli25fv06U6dOJUuWLGrHS5P+/vtvhg0bFvf9FiK1K1WqFKdPn6Z169Z06tSJdu3aERoaqnYswesbPUuVKiVLN1QkxbMQaYTBYGDlypV8+eWX9OrVi9q1a3PlyhV+++03cufOrXa8NOvy5cu0bt2aBg0aMGHCBLXjCGEydnZ2LFq0iNWrV7Njxw7c3d05c+aM2rEEr5du7N69Wz4RUIkUz0KkcoqisG3bNlxdXWnfvj2lSpXi3LlzLF++HCcnJ7XjpWkhISE0atSIPHnysHLlStmtRKRJrVu35uzZszg4OFCxYkVmzZolRZvK9Ho99+/f58KFC2pHSZfkN70Qqdi+ffuoUKECjRs3JmfOnJw6dYotW7ZQokQJtaOleQaDgdatW/P48WO2b99OxowZ1Y4kRJIpXLgwx48fp0+fPnz//fc0bNiQJ0+eqB0r3XrTqluWbqhDimchUqFTp05Ru3btuBtH9u3bx549eyhXrpzKydKPYcOGsWfPHtavX0+hQoXUjiNEkrO0tGTGjBn8+eefnDx5ktKlS3Po0CG1Y6VLNjY2VK1aVfZ7VokUz0KkIufPn6dx48ZUqFCBBw8esHXrVk6ePImnp6fa0dKVFStW8NNPPzFjxoy4NzBCpBf169fH39+fokWL4unpydixYzEYDGrHSnf0er206laJFM9CpALBwcFx65kDAwNZuXIlAQEBNGrUSDrYJTMfHx+6du3Kt99+S9++fdWOI4QqvvjiC/bu3cvYsWOZMGECnp6e3LlzR+1Y6YperyciIoLjx4+rHSXdkeJZiBTs3r179OzZE2dnZ/bt28dvv/3GpUuXaNu2LWZmZmrHS3fu3btH48aNcXd3Z968efLGRaRrZmZmjBo1igMHDhAcHEzp0qXZsWOH2rHSjZIlS5IjRw5ZuqECKZ6FSIGePn3KkCFDKFSoEOvWrWPSpElcu3aNHj16SMtnlURGRtKkSRO0Wi2bN2/GyspK7UhCpAjVqlXD39+fKlWq0LBhQ/r3709UVJTasdK8N626pXhOflI8C5GChIaGMmHCBJycnPjtt98YPHgwwcHBDB48GFtbW7XjpVuKotC9e3fOnTvH1q1byZkzp9qRhEhRsmbNytatW5k9ezbz5s2jUqVKBAUFqR0rzdPr9fj5+fH48WO1o6QrUjwLkQJERkYya9YsChUqxMSJE+ncuTPBwcGMHz8eBwcHteOlez///DPLly9n8eLFlC1bVu04QqRIGo2Gvn37cuLECUJDQ3F3d2flypVqx0rTpFW3OqR4FkJFsbGxLFq0iCJFijBo0CAaNWpEUFAQM2fOxNHRUe14AvD29mbw4MEMGTKENm3aqB1HiBTvTSfCJk2a0L59ezp27EhYWJjasdKkXLlyUbJkSVm6kcykeBZCBUajkXXr1uHi4kLXrl2pXLkyFy9e5Pfffydv3rxqxxP/CgoKolWrVtSpU4dJkyapHUeIVMPe3p7ly5ezbNkyNm7cSNmyZfH391c7VpokrbqTnxTPQiQjRVH4+++/KVOmDK1ataJo0aL4+fmxdu1aihYtqnY88R8vXrygYcOG5MiRg9WrV8vuJkIkwjfffMOZM2ewtramQoUKzJ07V4o8E9PpdNy7d49Lly6pHSXdkOJZiGRy5MgRqlWrRv369bG3t+fIkSP8+eefuLq6qh1N/A+DwUDbtm25f/8+27dvl3XnQnwGZ2dnTp48Sbdu3ejduzdNmzbl2bNnasdKM6pWrYqVlZUs3UhGUjwLkcTOnj1L3bp1qVatGq9evWLnzp0cOnSIKlWqqB1NfMDIkSPZuXOnfCIghIlYW1vzyy+/sHXrVg4dOoSrqyvHjh1TO1aaYGtrK626k5kUz0IkkcuXL+Pl5UWZMmW4ceMGGzZswNfXlzp16khzjRRszZo1TJkyhalTp1KnTh214wiRpjRq1Ah/f3/y589P9erVmThxorT2NgGdTsehQ4dkf+1kIsWzECZ269YtOnXqRPHixfHx8WHJkiWcP3+e5s2bo9XKj1xK8+OPP+Lt7Q3AmTNn6NSpE+3bt2fgwIEqJxMibcqXLx8HDhxg+PDhjBo1Cr1ez/3799WOlarp9XrCw8OlVXcy0Siycl8Ik3j48CGTJk1i/vz5ZMqUiZEjR9KtWzfpRJeCPXr0iBw5cqDRaBgxYgRLly7liy++4NChQ1hbW6sdT4g0b//+/bRt2xaDwcDy5cvl055EMhqN5MqVi06dOjF58mS146R5Mg0mxGcKCQlh5MiRFCpUiGXLljFmzBiuX79Onz59pHBO4U6cOAG83gXlxx9/5OnTp6xevVoKZyGSiaenJwEBAZQtW5a6desyePBgoqOj1Y6V6mi1WnQ6HXv27FE7SrogxbMQifTq1SumTJlCwYIFmTlzJn369OHGjRv88MMPZMiQQe14Ih5OnDiBubl53J+jo6Np3rw5d+7cUTGVEOmLo6Mjf/75Jz/99BOzZs2iatWqBAcHqx0r1dHpdJw9e1ZadScDKZ6FSKDo6Gjmzp1L4cKFGT16NG3btuX69etMnjyZzJkzqx1PABHGCEIMIbw0vCRWif3geYcPHyY29v+PG41G/P39GTt2bDKkFEK8odVqGThwIMeOHePJkye4ubmxbt06tWOlKjqdDkVR2Ldvn9pR0jxZ8yxEPBkMBlauXMnYsWO5ffs27du3Z8yYMRQsWFDtaOlerBLLlegrXIu+xgPDAyKVyLhjGjRk1mbmC/MvKG5VnBzmOYDXb4Ls7OziimeNRoO9vT29e/emX79+0h5dCJW8ePGC7t27s27dOrp27cqsWbOwtbVVO1aqULJkScqVK8fixYvVjpKmmX/6FCHSN0VR2LJlCyNHjuTSpUs0bdqUv/76CxcXF7WjpXtGxciZqDP4RvgSTTQaNCi8PR+goPDM+Izn0c8JjA7E0cyRmrY18d/vH1c4582bl6FDh9KxY0fs7OzUeCpCiH85ODiwZs0adDodffr04dixY6xbt44SJUqoHS3F0+l0bNiwAUVRZEvUJCTLNoT4AEVR2LNnD+XKlaNZs2bkzZsXHx8fNm3aJIVzCvDc8Jy1oWs5HnGcaF7fYPS/hfN/vTn22PCY9aHreVX0FYWKvL7J88aNG/Tq1UsKZyFSCI1GQ+fOnfH19UWr1eLh4cHChQultfcn6PV67ty5w+XLl9WOkqZJ8SzEe5w4cQJPT0/0ej3m5uYcOHAAb29vPDw81I4mgEexj1gXuo4nhicJvlb59587me8w6+ws2rZvi5mZWRKkFEJ8LhcXF3x8fOjQoQPdu3enZcuWhISEqB0rxapWrRqWlpay60YSk+JZiP84d+4cDRs2pFKlSjx9+pTt27dz/PhxatSooXY08a+XhpdsDttMtBL90Znm+Lgec5394ftNlEwIkRRsbGyYP38+69evZ/fu3bi5uXHq1Cm1Y6VItra2VKlSRVp1JzEpnoUArl27Rtu2bXF1deXixYusXr0af39/vv76a1k3loIoisKe8D0mKZzfuBh9kWvR10wylhAi6Xh5eeHn50eOHDmoUqUK06ZNw2g0qh0rxdHr9Rw4cEBadSchKZ5Funb37l26d+9OsWLFOHToEPPnz+fSpUu0bt1aWmmnQBejL3In9o7JCuc39oXvI0qRFxohUrqCBQty5MgRBg4cyNChQ6lbty4PHz5UO1aK8qZV95smUML0pDoQ6dKTJ08YNGgQhQoVYtOmTUydOpWgoCC6deuGhYWF2vHEeyiKwunI00kydqQSyeUoucFGiNTAwsKCKVOm4O3tjb+/P6VLl2bv3r1qx0oxSpcuTfbs2WXdcxKS4lmkKy9fvmTcuHE4OTmxYMEChg0bRnBwMAMHDsTGxkbteKlOgQIFcHR0JCYmJu5rBw4cQKPR0L9/fwC2b9/O999/D8DBgwdxdXUF4ObNm2TKlClejxMWFkaXfl0Y7DaYaVWnMb36dLaN2oYhxsCp1adY1G7RZz+XgKiAT97J37FjR2bNmvXJsbZu3crJkyc/O5MQ4sP0ej0BAQGULFkSvV7PDz/88NbvovRKq9VSu3ZtWfechKR4FulCREQEM2fOxMnJicmTJ9OtWzdu3LjB2LFjyZgxo9rxUrV8+fKxffv2uD8vXryYsmXLxv25YcOG/Pzzz4keX1EUGjRowKPQRww7OowhR4YwYO8AsjllIzbqw90DE+q58TlhSphJxpLiWYjkkTNnTry9vZk0aRLTpk2jRo0a3Lp1S+1YqtPr9Zw5c4anT5+qHSVNkuJZpGkxMTEsXLiQIkWKMGTIEJo1a8a1a9f46aefyJYtm9rx0oRvv/2WJUuWAK87g508eZI6derEHf/jjz9o3LjxR8eIjo6mXbt2dO/eHYPB8Nax/fv3v76h86e2mNu87utkZmFG5W8rY5XB6vX1r6JZ1mUZUytPZYbnDJ7c/P8t7Pb/sp8pFacwtfJUVnRbQcTLCABio2PZNnobUypNYVrVacxvPp9HsY8wGAwMHjyYEiVKUKJECfr06UN0dPQ7mfft20fFihVxc3OjePHicR29/v77b7Zv38706dNxdXVl0aLXs+IrVqygfPnyuLu7U61aNQICAhLybRZCfIBWq2XYsGEcPnyYO3fu4OrqyubNm9WOpSpp1Z20pHgWaZLRaGTNmjW4uLjQo0cPqlevzuXLl1mwYAF58uRRO16aUrlyZW7evMm9e/dYs2YNXl5eCdo3OSQkhDp16uDi4sKCBQveufbMmTOUKVOGl2YvPzjG7bO3aTCqAUOPDaVo9aLsm/36BePinoucXHWSvrv6MvTYUCxtLdkxbgcAe3/ey+Prjxl0YBBDjgyh/fz2PDM8Y+HChZw+fZozZ87g7+/P9evX3ztz7u7uztGjR/Hz8+PIkSOMHz+eO3fuUK9ePRo2bMjgwYPx9/enS5cuHDt2jDVr1nD48GHOnj3LxIkTadOmTby/R0KIT6tUqRL+/v54enrSrFkzevXqRWRkpNqxVPHFF1/g4uIiSzeSiBTPIk1RFIUdO3bg5uZGmzZt+PLLL/H392fVqlUULlxY7XhpVvv27fnjjz9YsmQJnTp1ivd10dHRVK5cmY4dO/LDDz989FyDYvjgsQIeBciaP2vcvz+9+fqjyquHruLWxA1bB1sAKneqzNWDVwG4sPsC1bpXw9zq9Wy2fTZ7Yoll7969dOzYESsrK8zNzenatet7b7x5+vQpXl5elChRAk9PT54+fcr58+ffm2/btm0EBARQvnx5XF1d6dOnD8+ePSMiIuIT3yEhREJkzpyZjRs38ttvv7F48WLKly/PpUuX1I6lCr1ez+7du6UrYxKQ4lmkGQcPHqRy5co0bNiQzJkzc+zYMbZv306pUqXUjpbmffPNN/zyyy9YW1tTpEiReF9nYWFBlSpV2LFjxwdv9ClTpgxnz57FGPPh/VzNrc3j/l1rpsUQ+4FC+yNbdisomPHujPmH9vnu0aMHVapUITAwEH9/f4oWLfrBWS5FUejQoQP+/v5x/7t//77cpCpEEtBoNHz33Xf4+PgQHR1N2bJlWbp0aborIvV6Pf/88w9Xr15VO0qaI8WzSPV8fX3R6/XUrFmTmJgYdu/ezYEDB6hUqZLa0dKN3LlzM3nyZKZOnZqg6zQaDQsWLCBv3rw0btz4vTOxnp6eFCxYkK3DthIT+brANsQaOP7HcaLCPr43c9HqRfHf6k/ky9dF7fE/juNc0xmAEnVKcHjB4bibDkOfhJLJLBO1a9dm+fLlREdHExsby6JFi9Dr9e+M/fz5c/Lnz49Go+Hw4cNvrWHOmDEjL168iPtzw4YNWblyJbdv3wZeLyvy9fVNyLdKCJFApUqVwtfXl1atWtGpUyfatWvHy5cfXv6V1rxp1S1LN0xPimeRal28eJFmzZrh4eHBnTt32LRpEz4+Puh0OukKqIJvv/2WihUrJuramTNn4u7uTr169QgLe3vHC41Gw19//UVG64xMqTTl9Q1+VabxOPjxWzPO7+Oic6F8m/LM+moWUytPJSo0igajGwBQq18tshfKzk81fmJatWms6rkKRzNHunXrhru7O+7u7ri6ulKgQIG4bff+a8qUKQwbNgxXV1eWLFlC+fLl4461b9+e9evX4+bmxqJFi6hatSrTpk2jSZMmlC5dmuLFi7N27dpEfa+EEPFnZ2fH4sWLWbVqFTt27MDd3T3dvHG1s7OjcuXKUjwnAY2S3j7HEKnezZs3GTt2LCtWrCBv3ryMGzeOdu3aJegmNZH6XI++zp+v/kyy8W01tnRx6CJvvIRIo65du0arVq04d+4cU6dOpX///mn+533KlClMnDiRp0+fYmlpqXacNENmnkWq8eDBA3r37k3RokXZtWsXv/zyC1euXKFDhw5SOKcDBS0KYquxTZKxNWgobVU6zb+QCpGeFS5cmOPHj9O7d28GDBjA119/zZMnTz59YSqm0+kICwuTfedNTIpnkeI9f/6c4cOH4+TkxKpVqxg/fjzXr1+nV69eWFlZqR1PJBOtRourtWvSjI2W4lbFk2RsIUTKYWlpycyZM/nzzz85efIkpUuX5tChQ2rHSjJubm5kzZpVlm6YmBTPIsUKCwtj0qRJFCxYkF9++YXvv/+e4OBghg0bhp2dndrxhArcrdzJos2C5mPbZiRCVZuq2Gnl75QQ6UX9+vUJCAigSJEieHp6MnbsWGJjTdexNKXQarXodDopnk1MimeR4kRFRfHLL79QqFAhxo0bR4cOHQgODmbixIlkzpxZ7XhCRWYaM76y+8pkxbMGDXnM81DKSrYzFCK9+eKLL9i3bx9jxoxhwoQJ1KpVizt37qgdy+R0Oh2+vr48e/ZM7ShphhTPIsWIjY1l6dKlFC1alO+//5769etz9epVZs+eTY4cOdSOJ1IIR3NH6tvVR/uZv740aMhmlo0GGRrIWmch0ikzMzNGjx7NgQMHuH79OqVLl2bHjh1qxzIpadVtelI8C9UZjUY2btxIyZIl6dSpE+XKlePChQssWbKE/Pnzqx1PpEBOlk40ztAYG41NomehC1oUpJl9M6w0sm5eiPSuWrVqBAQExDXa6t+/P1FRH99HPrXImzcvX375pSzdMCEpnoVqFEXB29sbDw8PvLy8KFCgAL6+vmzYsIFixYqpHU+kcHkt8vJNxm9wtnjd9ORTRfSb49Yaa+rY1aGBXQMpnIUQcbJmzcq2bduYPXs28+bNo2LFigQFBakdyyT0ej179uxJd10Wk4oUz0IVx44do0aNGtSpUwdra2sOHTrEzp07KVOmjNrRRCpirbXmqwxf0TFjR8pYl8Fea//e88wwI5d5Lr6y/YrODp1xtnSWpRpCiHdoNBr69u3LiRMnCAsLw93dnZUrV6od67PpdDpu3bqVZt4MqE2apIhk5e/vz4gRI/j7778pXbo0EydOpF69elLICJOJMkbxxPCEaKLRosVOa0cWbRa0GpkrEELEX2hoKD179mTlypV06NCBOXPmkCFDBrVjJUpYWBhZsmRh5syZ9O7dW+04qZ4UzyJZXL16lTFjxrB27VqKFCnChAkT8PLyQquVgkYIIUTKtXz5cnr27EmePHlYu3Ytrq6uakdKlJo1a5IxY0a2bdumdpRUTyoXkaT++ecfunbtiouLC0ePHuX333/nwoULtGzZUgpnIYQQKd4333zDmTNnsLa2pkKFCsydOzdVrh3W6XTs37+fmJgYtaOkelK9iCTx+PFjBgwYQJEiRdi6dSs//fQTQUFBdOnSBQsLC7XjCSGEEPHm7OzMyZMn6dq1K71796Zp06apbt9kvV4vrbpNRIpnYVIvXrxg9OjRODk5sXjxYkaMGEFwcDD9+/fH2tpa7XhCCCFEolhbW/Prr7+yZcsWDh06hKurK8eOHVM7Vry9adW9Z88etaOkelI8C5MIDw9n+vTpODk5MX36dL777juCg4MZNWoU9vbv3wFBCCGESG0aN26Mv78/+fLlo3r16kycOBGDwaB2rE8yMzOjVq1ast+zCUjxLD5LTEwM8+fPp3Dhwvzwww+0aNGC69evM23aNLJmzap2PCGEEMLk8uXLx8GDBxk+fDijRo1Cr9dz//59tWN9kl6v5/Tp0zx//lztKKmaFM8iUQwGA6tWraJYsWL07NmTWrVqcfnyZebNm0fu3LnVjieEEEIkKXNzcyZMmMDevXu5ePEipUuXZufOnWrH+iidTofRaGT//v1qR0nVpHgWCaIoCtu2bcPV1ZV27dpRsmRJAgICWLFiBYUKFVI7nhBCCJGsPD09CQgIoGzZstSrV4/BgwcTHR2tdqz3ypcvH87OzrJ04zNJ8Szibf/+/VSsWJHGjRvj6OjIyZMn2bp1KyVLllQ7mhBCCKEaR0dH/vzzT3766SdmzZpF1apVCQ4OVjvWe+n1ery9vVPldnsphRTP4pNOnTpF7dq1qVWrFoqisHfvXvbt20f58uXVjiaEEEKkCFqtloEDB3Ls2DGePHmCm5sb69atUzvWO/R6Pbdu3eLatWtqR0m1pHgWH3T+/HmaNGlChQoVePDgAVu2bOHkyZPUqlVL7WhCCCFEilSuXDnOnj1L3bp1adWqFd26dSM8PFztWHFq1KiBhYWFbFn3GaR4Fu8IDg7mm2++oVSpUgQEBLB8+XICAgJo3LgxGo1G7XhCCCFEiubg4MCaNWtYtGgRK1euxMPDg/Pnz6sdC4AMGTJQsWJFWff8GaR4FnHu379Pz549cXZ2Zu/evcydO5fLly/Tvn17zMzM1I4nhBBCpBoajYbOnTvj6+uLVqvFw8ODhQsXpoi1xnq9Xlp1fwYpngVPnz5l6NChFCpUiLVr1zJx4kSuXbvGd999h6WlpdrxhBBCiFTLxcUFHx8fOnToQPfu3WnZsiUhISGqZtLr9YSGhuLj46NqjtRKiud0LDQ0lB9//BEnJyfmzp3LwIEDuXHjBkOGDMHW1lbteEIIIUSaYGNjw/z581m/fj27d+/Gzc2NU6dOqZbH3d2dzJkzy9KNRJLiOR2KjIxk1qxZFCpUiAkTJtCpUyeCg4OZMGECDg4OascTQggh0iQvLy/8/PzIkSMHVapUYdq0aRiNxmTPYWZmRu3ataV4TiQpntOR2NhYFi9eTNGiRRk4cCANGzYkKCiIn3/+GUdHR7XjCSGEEGlewYIFOXLkCAMHDmTo0KHUq1ePhw8fJnsOvV6Pj4+P6ktIUiMpntMBo9HI+vXrKV68OF26dKFixYpcvHiRRYsWkS9fPrXjCSGEEOmKhYUFU6ZMwdvbGz8/P1xdXdm7d2+yZpBW3YknxXMapigKf//9N2XKlKFly5YULlyYs2fPsm7dOpydndWOJ4QQQqRrer2egIAASpQogV6vZ8SIEcm2A0b+/PkpWrSoLN1IBCme06gjR45QrVo16tevj729PUeOHOGvv/7Czc1N7WhCCCGE+FfOnDnx9vZm0qRJTJ06lRo1anDr1q1keWy9Xi/NUhJBiuc05k1Xo2rVqvHq1St27tzJoUOHqFKlitrRhBBCCPEeWq2WYcOGcfjwYe7evYurqyubN29O8sfV6XQEBwdz/fr1JH+stESK5zTiypUrtGjRgjJlynDjxg3Wr1+Pr68vderUka6AQgghRCpQqVIl/Pz88PT0pFmzZvTq1YvIyMgke7waNWpgbm4uSzcSSIrnVO727dt07twZFxcXTp48yeLFizl//jxeXl5otfKfVwghhEhNMmfOzMaNG/ntt99YvHgx5cuX5/Lly0nyWBkzZqRixYqydCOBpLpKpR49ekT//v0pUqQIO3bs4OeffyYoKIhOnTphbm6udjwhhBBCJJJGo+G7777Dx8eH6OhoypQpw9KlS5Oktbder2ffvn3ExsaafOy0SornVCYkJISRI0fi5OTEH3/8wejRowkODqZv375YWVmpHU8IIYQQJlKqVCl8fX1p1aoVnTp1ol27doSGhpr0MXQ6HS9fvpRW3QkgxXMqER4eztSpU3FycmLmzJn07t2b4OBgRowYQYYMGdSOJ4QQQogkYGdnx+LFi1m1ahU7duzA3d2dM2fOmGz8smXLkilTJln3nABSPKdw0dHRzJ07l0KFCjFy5Ehat27N9evXmTJlClmyZFE7nhBCCCGSQZs2bTh79iwODg5UrFiRWbNmmWQZx5tW3bLuOf6keE6hDAYDy5cvx9nZmT59+qDX67ly5Qpz584lV65cascTQgghRDIrXLgwx48fp0+fPnz//fc0bNiQJ0+efPa4Op2OU6dOSavueJLiOYVRFIUtW7ZQqlQpOnTogJubG4GBgSxbtgwnJye14wkhhBBCRZaWlsyYMYM///yTEydOULp0aQ4dOvRZY+p0OgwGAwcOHDBRyrRNiucUQlEU9u7dS/ny5WnatClffPEFPj4+bN68meLFi6sdTwghhBApSP369QkICKBo0aJ4enoyduzYRO+YUbBgQYoUKSJLN+JJiucU4OTJk9SqVQudToeZmRn79+9n9+7deHh4qB1NCCGEECnUF198wd69exk7diwTJkygVq1a3LlzJ1Fj6XQ6uWkwnqR4VlFgYCCNGjWiYsWKPH78mG3btnH8+HFq1qypdjQhhBBCpAJmZmaMGjWKAwcOEBwcTOnSpdmxY0eCx9Hr9Vy/fp3g4OAkSJm2SPH8HjEGIxExBiJjDRiTYEPy69ev07ZtW0qXLs2FCxdYtWoV/v7+NGzYUFppCyGEECLBqlWrhr+/P1WqVKFhw4b079+fqKioeF9fs2ZNzMzM2LNnD0bFyCvjK8KMYUQp8R8jvdAoSdGuJpUxGBXuhkZwNyyS5xExRBqMcce0GshoaUF2W0sKZLLF3jLx3fvu3r3LhAkTWLx4MY6OjowePZpOnTphYWFhiqchhBBCiHROURR+/fVXBg8eTIkSJVi7di1FihT55HXPDc8Z8ccIcrvlJlPBTBgwxB2z09iR0zwnhS0KU9iyMOaa9N3JOF0Xz0ZFIejZK64+CyPG+PFvgwZQgOy2lpR2zEhGq/gXvE+ePGHq1KnMmTMHW1tbhg8fTq9evbCxsfm8JyCEEEII8R5nz56lVatW3L9/n3nz5tGuXbv3nvfM8IwD4Qe4E3vndaHzgQ/ANWhQULDSWFHWuizuVu5oNelzAUO6LZ5Do2PxufecF1EJuzP1zd+p4tntKZLZ7qPLLEJDQ5k5cyYzZsxAURQGDhzIgAEDyJgx42ckF0IIIYT4tNDQUHr27MnKlSvp0KEDc+bMietKrCgKZ6POcjziOMq//yREdrPs1LWrS2azzEkRPUVLl8Xz88gYjvzzFIMxoX9V3pY/ow3uOR3eKaAjIiKYN28ekydPJjQ0lF69ejFs2DCyZ8/+ecGFEEIIIRJo+fLl9OzZkzx58rB27VpKly7NvvB9XIi+kOgxNWiwwIKm9k3JYZ7DhGlTvnRXPL+KjmX/rSfEfmbh/EaRzHaUdHw9kxwTE8Mff/zBuHHjePDgAZ06dWLUqFHkzZvXBI8khBBCCJE4V65coWXLlly+fJlp+6ahuHx+FaRBg4XGgjb2bXAwczBBytQhXS1WURSF0w9CTFY4AwQ9f8XDsEjWrFmDi4sL3bp1o1q1aly6dImFCxdK4SyEEEII1Tk7O3Py5En6Te5nksIZQEEhRonB+5U36WkuNl0VzzdCwnkWEWOywvmNvwOu8E2HjhQrVgx/f39Wr14drztbhRBCCCGSi7mVOUW/LYpGMd22uAoK9w33CYwONNmYKV262WtEURSuPAtLkrEdsjmy8/gpapd1TZLxhRBCCCE+17Xoa4QpYR/cUeNz+Eb4UtKyZLroV5GmZp4LFCiAs7Mzrq6uuLi4MHfu3LhjD19FERFr/MjV7/I9sIfR7ZvF61yz7Hne+/WxY8cSGRmZoMcVQgghhPhc/62LXF1dqeNRh/sX779z3ojCI3h6+ykAC1os4GHQQwB+/fpXzv117r1j/++xUCWUW7G3kuBZpDxpqngGWLduHf7+/uzcuZMffviBc+de/4d98CoKxWjEaExYAR1fzyNjiDG8O/a4ceOkeBZCCCFEklu3bh1169blwIEDcWuQ39RFp86e4vtD35PLJddHx+i+vjs5iiR89wwtWm7G3ExM7FQnzS7byJ8/P87OzrRp0wZnZ2duP37Gg3t3Gb1kLYEnj7Ft8TwAsuXKTY/x08iaIxexMTEsmTSagOOHyZDRgS/LlI8b7/yp4yydPJoZW/cCcPvqZSb1+Ib5+30AOHNwLxPazMYYG4NGo2HBggUsXboUgKpVq2JmZsbu3btxdHRM5u+EEEIIIdKDo0ePsmvXLnbt2kXlypXfmrx7ZHgU9+/nd55n+9jtmFmY8WWtL98aY1zpcXRe2Zk8JV9/oh50OIg9M/cQERJBibolaDSh0TtLMyJDI9k6ciuPLzzGNsaWChUqMGfOHCwtLZPw2aonzRbPgYGBXL58ma+//poDBw4wZaM3GbJm4/bVy6yYPoFpm3aRNUcuNs6fzbyRgxj5+yr2rF/JvRvXmbXjAAATurSJ12Pdu3GdOT98z7KNW6lcwpmYmBgiIiKYMGECCxYsYMuWLTg4vN7C5fHjx0n2nIUQQgiRfkVERGBhYUFMTAwnT57EYDBQsWJFcuXKhYW9BWHGMLqs6sLq3qvp+1dfchbLyfE/jvPq2asPjvngygP67+qPIcbArw1+5eyms5RpXuatc7aN2oZTRSfaz27Pd5m+o2vXrsyePZvBgwcn9VNWRZornlu2bImNjQ22trYsWbKE8+fPY2Njg33WbCjA+VPHcK1Sk6w5Xn9sUad1BzbM/RmDwUDgyaNUb9Qci3/fKXk2a8X+jWs++ZgBxw/jWqUGf6xcTZMl8945LjtvCCGEECI5GQwGACIjI7lx4wZzD80ltmQsgTsDyV08NzmL5QSgQvsKbBq26YPjeLT0wMzCDDMLM8p4leHKwSvvFM+Bfwdy8/RNDv12iIVmC4mIiMDMzCzpnpzK0lzxvG7dOlxdXeP+fP78eTJkyIBWA4b37FH3sbtC/3vMzNwM43/WNEdHRf3PudCxYwc6Narz1tcbNWrEqlWr4tphCiGEEEIkhYULF7Jr1664whnA2tqaVq1a4eHuwYmYE+9elMDNMd5XNymKwrfLviV34dz0ytwrobFTnTRXPH+IrYU5odGxlChfmU0LfuHZwwdkyZET77XLKVmxCmZmZpSqWJXDOzZRtUETFEXhwOZ1cdfnyJOfx/fu8OLZUxyyZOXQ9o1xx1yr1GDDbz+TZdQoqlVtSExMDOHh4Tg4OGBvb0/lypXJnz+/Gk9bCCGEEOnEnj17MBqNaLVa7O3t0Wg07NixgypVqhAcHYwSo1DAowD3et/j4dWH5Ciag1MrT2GINnxwTN/1vpRpXgZDrIGzm85S/bvq75xTsl5J9s3eR4/ZPQB4/vw5T58+pXDhwkn2XNWUbornLNYWhEXHkq9oMdoPHsWErm2B1zcMfjdhOgC1vdpyO+gK/RrUiLthMPjC6906suTISeMuPRnmVQ+HbNlxr1ozbuxc+QvSa9LP9O/WGUNsDGZmZsyfP59y5coxcOBAdDodtra2csOgEEIIIZLMF198QcaMGRkyZAi9e/emVKlSdO7cGRsbG4wYeWp4SpOJTWj9a2sWt1+MuaU5xWoVwy6L3QfHzOGcg9l1ZxP+PJwSdUvg3tT9nXOaTGzCn+P/ZGzVscw0n4m5uTnTpk1Ls8WzRkkn/RRvvQjnzIMXSTa+rYUZdZykMBZCCCGEehRF+eCS1KUvlvLS+DLJHltvq+dLqy8/fWIql+b2ef6QL+xtMNcmXdebQplsk2xsIYQQQoj4+Ni9XKWtSifZ41piSRHL9LFBQropns21GpySqMA102jI7yDFsxBCCCFSLhdLF8yTYMWuBg2lrEthrkkfq4HTTfEMUCxrBmzMTf+USzlmxNIsXX0rhRBCCJHKWGutqWZbzaRjatCQQZOBctblTDpuSpauKj5zrZayuTKZbDwN4GhrSQEHG5ONKYQQQgiRVEpYliCfeT40Cd2j7iP0dnosNBYmGy+lS1fFM0B2WyvK5nT47HE0gIOVBeW/yPzR9UVCCCGEECmFRqOhXoZ6ZDfLbpICWm+rJ49FHhMkSz3SzW4b/+teaCS+D0IwGBUS8w3IlcGKsrkyYaFNd+8/hBBCCJHKRSvReL/yJjgmOMHXatBggQV6Oz2FLAslQbqULd0WzwARsQb8HrzgwasoNBCvItpCq8E1hwN57K1lxlkIIYQQqZaiKFyJucLB8INEKVFo0PCxKcU3x50snPC09cRO++H9odOydF08v/EiKobgkHDuhkYS/Z8W3G9ogEzWFjhlsiWPvQ1mSbjlnRBCCCFEcopVYrkafZXAqEAeGh6+t4C20dhQ2LIwpaxKkc0smwopUw4pnv9HRKyBl1GxxBqNaNBgY2GGg5U5WpllFkIIIUQaZ1AMPDU8JdQYihEjlhpLspllS7ezzO8jxbMQQgghhBDxJHe7CSGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT1I8CyGEEEIIEU9SPAshhBBCCBFPUjwLIYQQQggRT/8Hyx2mF8oBS4UAAAAASUVORK5CYII=",
            "text/plain": [
              "<Figure size 700x500 with 1 Axes>"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        }
      ],
      "source": [
        "input_text = (\n",
        "    \"Some products are edible and others are inedible. Soap, newspapers, and shoes, for example, \"\n",
        "    \"are inedible. Of the products that are edible, some are sweet and others are savory. \"\n",
        "    \"Chocolate, candy, and ice cream are sweet, while pizza, burgers, and fries are savory. \"\n",
        "    \"Chocolate comes in different forms, such as milk chocolate, dark chocolate, and white chocolate. \"\n",
        "    \"The New York Times, The Wall Street Journal, and The Washington Post are newspapers.\"\n",
        ")\n",
        "\n",
        "class Node(BaseModel):\n",
        "    id: int\n",
        "    label: str\n",
        "    color: str\n",
        "\n",
        "class Edge(BaseModel):\n",
        "    source: int\n",
        "    target: int\n",
        "    label: str\n",
        "    color: str = \"black\"\n",
        "\n",
        "class KnowledgeGraph(BaseModel):\n",
        "    nodes: List[Node] = Field(..., default_factory=list)\n",
        "    edges: List[Edge] = Field(..., default_factory=list)\n",
        "\n",
        "knowledge_graph = llm(\n",
        "    KnowledgeGraph,\n",
        "    \"Format the information in the user-provided text as a knowledge graph.\",\n",
        "    input_text,\n",
        ")\n",
        "\n",
        "visualize_graph(knowledge_graph)\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "Some structures are recursive. Consider the task of parsing a linguistic sentence into a grammatical tree structure. This NLP task has kept computational linguists busy for decades, often with limited success. LLMs, however, are quite good at this sort of thing. Let's try to get the LLM to parse a simple sentence into a simplified tree grammar of English."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 10,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "NOUN:\n",
            "   DET:\n",
            "      the\n",
            "   ADJ:\n",
            "      quick\n",
            "      brown\n",
            "   NOUN:\n",
            "      fox\n",
            "VERB:\n",
            "   jumps\n",
            "PREP:\n",
            "   PREP:\n",
            "      over\n",
            "   NOUN:\n",
            "      DET:\n",
            "         the\n",
            "      ADJ:\n",
            "         lazy\n",
            "      NOUN:\n",
            "         dog\n"
          ]
        }
      ],
      "source": [
        "input_str = \"the quick brown fox jumps over the lazy dog\"\n",
        "\n",
        "class NounPhrase(BaseModel):\n",
        "    det: str\n",
        "    adj: List[str]\n",
        "    noun: str\n",
        "\n",
        "class PrepPhrase(BaseModel):\n",
        "    prep: str\n",
        "    noun: NounPhrase\n",
        "\n",
        "class VerbPhrase(BaseModel):\n",
        "    noun: NounPhrase\n",
        "    verb: str\n",
        "    prep: PrepPhrase\n",
        "\n",
        "grammar_tree = llm(\n",
        "    VerbPhrase,\n",
        "    \"Parse the user-provided sentence into a simple grammar tree.\",\n",
        "    input_str,\n",
        ")\n",
        "\n",
        "print_tree(grammar_tree.dict())\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "💡 What other cases do you know where parsing unstructured information from text into data structures is a useful application? Could you automate any process that is currently performed manually by giving the task to an LLM? ... Some models (like GPT-4o) can also read images as input ... how about using a photo or scan as the input and parsing it into structured data?"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Synthetic Data Generation\n",
        "\n",
        "They used to say that \"data is the new oil\". That valuable! What if we found an endless supply of data to work with? LLMs are great at generating new texts and pieces of information. That can be very useful in many data science and ML projects, as we can use the LLM to generate synthetic data for us.\n",
        "\n",
        "Let's try using the LLM to generate some test data for exercising a sentiment analysis system."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 16,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "{\n",
            "  \"sentiment\": \"The pasta was incredibly bland and overcooked. The sauce tasted like it came straight out of a can. I was really disappointed.\",\n",
            "  \"rating\": 1\n",
            "}\n"
          ]
        }
      ],
      "source": [
        "class SyntheticSentiment(BaseModel):\n",
        "    sentiment: str = Field(..., description=\"A review about food.\")\n",
        "    rating: int\n",
        "\n",
        "sentiment = llm(\n",
        "    SyntheticSentiment,\n",
        "    \"Generate food review with sentiments within a spectrum of sentiments, with rating between 1 and 5.\",\n",
        "    temperature= 0.5\n",
        ")\n",
        "\n",
        "print_result(sentiment)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "That's easy. Now let's generate multiple examples."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 17,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_be43e_row0_col0, #T_be43e_row0_col1, #T_be43e_row1_col0, #T_be43e_row1_col1, #T_be43e_row2_col0, #T_be43e_row2_col1, #T_be43e_row3_col0, #T_be43e_row3_col1, #T_be43e_row4_col0, #T_be43e_row4_col1, #T_be43e_row5_col0, #T_be43e_row5_col1, #T_be43e_row6_col0, #T_be43e_row6_col1, #T_be43e_row7_col0, #T_be43e_row7_col1, #T_be43e_row8_col0, #T_be43e_row8_col1, #T_be43e_row9_col0, #T_be43e_row9_col1 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_be43e\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_be43e_level0_col0\" class=\"col_heading level0 col0\" >sentiment</th>\n",
              "      <th id=\"T_be43e_level0_col1\" class=\"col_heading level0 col1\" >rating</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_be43e_row0_col0\" class=\"data row0 col0\" >The food was completely inedible. There was nothing good about it, and I felt sick afterward.</td>\n",
              "      <td id=\"T_be43e_row0_col1\" class=\"data row0 col1\" >*</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_be43e_row1_col0\" class=\"data row1 col0\" >Highly disappointing. The flavors were off, and the presentation was poor.</td>\n",
              "      <td id=\"T_be43e_row1_col1\" class=\"data row1 col1\" >*</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_be43e_row2_col0\" class=\"data row2 col0\" >Not worth the money. The portions were small, and the taste was bland.</td>\n",
              "      <td id=\"T_be43e_row2_col1\" class=\"data row2 col1\" >**</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_be43e_row3_col0\" class=\"data row3 col0\" >It was okay. Nothing special, but not terrible either.</td>\n",
              "      <td id=\"T_be43e_row3_col1\" class=\"data row3 col1\" >**</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_be43e_row4_col0\" class=\"data row4 col0\" >Decent meal, but I’ve had better. The service was good though.</td>\n",
              "      <td id=\"T_be43e_row4_col1\" class=\"data row4 col1\" >***</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row5\" class=\"row_heading level0 row5\" >5</th>\n",
              "      <td id=\"T_be43e_row5_col0\" class=\"data row5 col0\" >A pleasant experience. The food was tasty, but a bit overpriced.</td>\n",
              "      <td id=\"T_be43e_row5_col1\" class=\"data row5 col1\" >***</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row6\" class=\"row_heading level0 row6\" >6</th>\n",
              "      <td id=\"T_be43e_row6_col0\" class=\"data row6 col0\" >The food was quite good, and the ambiance was nice.</td>\n",
              "      <td id=\"T_be43e_row6_col1\" class=\"data row6 col1\" >****</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row7\" class=\"row_heading level0 row7\" >7</th>\n",
              "      <td id=\"T_be43e_row7_col0\" class=\"data row7 col0\" >Very enjoyable meal. The flavors were excellent, and the staff was friendly.</td>\n",
              "      <td id=\"T_be43e_row7_col1\" class=\"data row7 col1\" >****</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row8\" class=\"row_heading level0 row8\" >8</th>\n",
              "      <td id=\"T_be43e_row8_col0\" class=\"data row8 col0\" >Absolutely delicious! Everything was cooked to perfection.</td>\n",
              "      <td id=\"T_be43e_row8_col1\" class=\"data row8 col1\" >*****</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_be43e_level0_row9\" class=\"row_heading level0 row9\" >9</th>\n",
              "      <td id=\"T_be43e_row9_col0\" class=\"data row9 col0\" >An unforgettable dining experience. The food was exquisite, and the service was impeccable.</td>\n",
              "      <td id=\"T_be43e_row9_col1\" class=\"data row9 col1\" >*****</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "                                                                                       sentiment rating\n",
              "0  The food was completely inedible. There was nothing good about it, and I felt sick afterward.      *\n",
              "1                     Highly disappointing. The flavors were off, and the presentation was poor.      *\n",
              "2                         Not worth the money. The portions were small, and the taste was bland.     **\n",
              "3                                         It was okay. Nothing special, but not terrible either.     **\n",
              "4                                 Decent meal, but I’ve had better. The service was good though.    ***\n",
              "5                               A pleasant experience. The food was tasty, but a bit overpriced.    ***\n",
              "6                                            The food was quite good, and the ambiance was nice.   ****\n",
              "7                   Very enjoyable meal. The flavors were excellent, and the staff was friendly.   ****\n",
              "8                                     Absolutely delicious! Everything was cooked to perfection.  *****\n",
              "9    An unforgettable dining experience. The food was exquisite, and the service was impeccable.  *****"
            ]
          },
          "execution_count": 17,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "class Rating(str, Enum):\n",
        "    poor = \"*\"\n",
        "    average = \"**\"\n",
        "    good = \"***\"\n",
        "    great = \"****\"\n",
        "    outstanding = \"*****\"\n",
        "\n",
        "class SyntheticSentiment(BaseModel):\n",
        "    sentiment: str = Field(..., description=\"A generated review about food.\")\n",
        "    rating: str = Field(...,description=\"rating of food review between 1 and 5\")\n",
        "\n",
        "class SyntheticSentiments(BaseModel):\n",
        "    sentiments: List[SyntheticSentiment]\n",
        "\n",
        "n = 10\n",
        "\n",
        "sentiments = llm(\n",
        "    SyntheticSentiments,\n",
        "    (f\"Generate {n} food review with sentiments within a spectrum of sentiments in \" + \n",
        "     \"sorted order from most negative to positive, with rating: \" +\n",
        "     json.dumps([rating for rating in Rating])),\n",
        "    temperature=0.7,\n",
        ")\n",
        "\n",
        "pd.DataFrame(sentiments.dict()[\"sentiments\"])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "💡 Where can you use unlimited amount of test data, conforming to strict ranges and structures? What projects or tests can you run if collecting a data set is as trivial as calling an LLM?"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KYFPu0w7kQry"
      },
      "source": [
        "## Decision Making\n",
        "\n",
        "We've looked at LLMs reading and, in a way, \"understanding\" information, and rewriting it in useful formats. But the best LLMs also exhibit limited, but nevertheless impressive, reasoning and decision-making capabilities. Let's see how we can exploit them."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Sentiment Analysis\n",
        "\n",
        "Sentiment analysis, passing a judgement on the tone of a linguistic statement, is a common task that is being used in many systems, especially ones that are user-facing. Without any additional training, our LLM turns out to be quite good at making these judgements. It even can judge its own confidence level."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 11,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_981f9_row0_col0, #T_981f9_row0_col1, #T_981f9_row0_col2, #T_981f9_row1_col0, #T_981f9_row1_col1, #T_981f9_row1_col2, #T_981f9_row2_col0, #T_981f9_row2_col1, #T_981f9_row2_col2, #T_981f9_row3_col0, #T_981f9_row3_col1, #T_981f9_row3_col2, #T_981f9_row4_col0, #T_981f9_row4_col1, #T_981f9_row4_col2, #T_981f9_row5_col0, #T_981f9_row5_col1, #T_981f9_row5_col2, #T_981f9_row6_col0, #T_981f9_row6_col1, #T_981f9_row6_col2 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_981f9\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_981f9_level0_col0\" class=\"col_heading level0 col0\" >text</th>\n",
              "      <th id=\"T_981f9_level0_col1\" class=\"col_heading level0 col1\" >sentiment</th>\n",
              "      <th id=\"T_981f9_level0_col2\" class=\"col_heading level0 col2\" >confidence</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_981f9_row0_col0\" class=\"data row0 col0\" >I am very happy with the service provided by the company.</td>\n",
              "      <td id=\"T_981f9_row0_col1\" class=\"data row0 col1\" >positive</td>\n",
              "      <td id=\"T_981f9_row0_col2\" class=\"data row0 col2\" >95%</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_981f9_row1_col0\" class=\"data row1 col0\" >The food was terrible and the service was slow.</td>\n",
              "      <td id=\"T_981f9_row1_col1\" class=\"data row1 col1\" >negative</td>\n",
              "      <td id=\"T_981f9_row1_col2\" class=\"data row1 col2\" >95%</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_981f9_row2_col0\" class=\"data row2 col0\" >The movie was okay.</td>\n",
              "      <td id=\"T_981f9_row2_col1\" class=\"data row2 col1\" >neutral</td>\n",
              "      <td id=\"T_981f9_row2_col2\" class=\"data row2 col2\" >80%</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_981f9_row3_col0\" class=\"data row3 col0\" >The weather is perfect for a day at the beach.</td>\n",
              "      <td id=\"T_981f9_row3_col1\" class=\"data row3 col1\" >positive</td>\n",
              "      <td id=\"T_981f9_row3_col2\" class=\"data row3 col2\" >95%</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_981f9_row4_col0\" class=\"data row4 col0\" >I am mostly satisfied with the product, but there are a few issues.</td>\n",
              "      <td id=\"T_981f9_row4_col1\" class=\"data row4 col1\" >neutral</td>\n",
              "      <td id=\"T_981f9_row4_col2\" class=\"data row4 col2\" >80%</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row5\" class=\"row_heading level0 row5\" >5</th>\n",
              "      <td id=\"T_981f9_row5_col0\" class=\"data row5 col0\" >The experience was note quite what I have expected.</td>\n",
              "      <td id=\"T_981f9_row5_col1\" class=\"data row5 col1\" >negative</td>\n",
              "      <td id=\"T_981f9_row5_col2\" class=\"data row5 col2\" >80%</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_981f9_level0_row6\" class=\"row_heading level0 row6\" >6</th>\n",
              "      <td id=\"T_981f9_row6_col0\" class=\"data row6 col0\" >Butterflies are often colourful, and they can fly.</td>\n",
              "      <td id=\"T_981f9_row6_col1\" class=\"data row6 col1\" >neutral</td>\n",
              "      <td id=\"T_981f9_row6_col2\" class=\"data row6 col2\" >95%</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "                                                                  text sentiment confidence\n",
              "0            I am very happy with the service provided by the company.  positive        95%\n",
              "1                      The food was terrible and the service was slow.  negative        95%\n",
              "2                                                  The movie was okay.   neutral        80%\n",
              "3                       The weather is perfect for a day at the beach.  positive        95%\n",
              "4  I am mostly satisfied with the product, but there are a few issues.   neutral        80%\n",
              "5                  The experience was note quite what I have expected.  negative        80%\n",
              "6                   Butterflies are often colourful, and they can fly.   neutral        95%"
            ]
          },
          "execution_count": 11,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "example_texts = [\n",
        "    \"I am very happy with the service provided by the company.\",\n",
        "    \"The food was terrible and the service was slow.\",\n",
        "    \"The movie was okay.\",\n",
        "    \"The weather is perfect for a day at the beach.\",\n",
        "    \"I am mostly satisfied with the product, but there are a few issues.\",\n",
        "    \"The experience was note quite what I have expected.\",\n",
        "    \"Butterflies are often colourful, and they can fly.\",\n",
        "]\n",
        "\n",
        "class Sentiment(str, Enum):\n",
        "    positive = \"positive\"\n",
        "    negative = \"negative\"\n",
        "    neutral = \"neutral\"\n",
        "\n",
        "class SentimentAnalysis(BaseModel):\n",
        "    sentiment: Sentiment\n",
        "    confidence: float\n",
        "\n",
        "results = []\n",
        "for example_text in example_texts:\n",
        "    sentiment_analysis = llm(\n",
        "        SentimentAnalysis,\n",
        "        \"Analyze the sentiment of the user-provided text.\",\n",
        "        example_text,\n",
        "    )\n",
        "    results.append({\n",
        "        \"text\": example_text,\n",
        "        \"sentiment\": sentiment_analysis.sentiment.value,\n",
        "        \"confidence\": f\"{sentiment_analysis.confidence * 100:.0f}%\",\n",
        "    })\n",
        "\n",
        "pd.DataFrame(results)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Classification\n",
        "\n",
        "Classification is another task that requires judgement. We want our LLM to take several pieces of content and assign them to a class, or multiple tags. We want the LLM to take out taxonomy into consideration, but also make a descision as to which tags would best fit every item."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 12,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_73b2f_row0_col0, #T_73b2f_row0_col1, #T_73b2f_row0_col2, #T_73b2f_row1_col0, #T_73b2f_row1_col1, #T_73b2f_row1_col2, #T_73b2f_row2_col0, #T_73b2f_row2_col1, #T_73b2f_row2_col2, #T_73b2f_row3_col0, #T_73b2f_row3_col1, #T_73b2f_row3_col2, #T_73b2f_row4_col0, #T_73b2f_row4_col1, #T_73b2f_row4_col2, #T_73b2f_row5_col0, #T_73b2f_row5_col1, #T_73b2f_row5_col2, #T_73b2f_row6_col0, #T_73b2f_row6_col1, #T_73b2f_row6_col2, #T_73b2f_row7_col0, #T_73b2f_row7_col1, #T_73b2f_row7_col2, #T_73b2f_row8_col0, #T_73b2f_row8_col1, #T_73b2f_row8_col2, #T_73b2f_row9_col0, #T_73b2f_row9_col1, #T_73b2f_row9_col2, #T_73b2f_row10_col0, #T_73b2f_row10_col1, #T_73b2f_row10_col2, #T_73b2f_row11_col0, #T_73b2f_row11_col1, #T_73b2f_row11_col2, #T_73b2f_row12_col0, #T_73b2f_row12_col1, #T_73b2f_row12_col2 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_73b2f\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_73b2f_level0_col0\" class=\"col_heading level0 col0\" >title</th>\n",
              "      <th id=\"T_73b2f_level0_col1\" class=\"col_heading level0 col1\" >subtitle</th>\n",
              "      <th id=\"T_73b2f_level0_col2\" class=\"col_heading level0 col2\" >tags</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_73b2f_row0_col0\" class=\"data row0 col0\" >The Great Gatsby</td>\n",
              "      <td id=\"T_73b2f_row0_col1\" class=\"data row0 col1\" >A novel by F. Scott Fitzgerald</td>\n",
              "      <td id=\"T_73b2f_row0_col2\" class=\"data row0 col2\" >literature</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_73b2f_row1_col0\" class=\"data row1 col0\" >The Theory of Relativity</td>\n",
              "      <td id=\"T_73b2f_row1_col1\" class=\"data row1 col1\" >A scientific theory by Albert Einstein</td>\n",
              "      <td id=\"T_73b2f_row1_col2\" class=\"data row1 col2\" >science physics</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_73b2f_row2_col0\" class=\"data row2 col0\" >The Technology and Culture of Ancient Rome</td>\n",
              "      <td id=\"T_73b2f_row2_col1\" class=\"data row2 col1\" >A cross-disciplinary study of ancient Rome</td>\n",
              "      <td id=\"T_73b2f_row2_col2\" class=\"data row2 col2\" >technology culture history</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_73b2f_row3_col0\" class=\"data row3 col0\" >Football on Television</td>\n",
              "      <td id=\"T_73b2f_row3_col1\" class=\"data row3 col1\" >The technology and cultural impact of televising football games</td>\n",
              "      <td id=\"T_73b2f_row3_col2\" class=\"data row3 col2\" >technology culture sports</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_73b2f_row4_col0\" class=\"data row4 col0\" >The Philosophy of Taylor Swift</td>\n",
              "      <td id=\"T_73b2f_row4_col1\" class=\"data row4 col1\" >A philosophical analysis of the music and lyrics of Taylor Swift</td>\n",
              "      <td id=\"T_73b2f_row4_col2\" class=\"data row4 col2\" >philosophy music</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row5\" class=\"row_heading level0 row5\" >5</th>\n",
              "      <td id=\"T_73b2f_row5_col0\" class=\"data row5 col0\" >The Spanish Language in popular music</td>\n",
              "      <td id=\"T_73b2f_row5_col1\" class=\"data row5 col1\" >A review of the use of the Spanish language in popular music</td>\n",
              "      <td id=\"T_73b2f_row5_col2\" class=\"data row5 col2\" >language music</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row6\" class=\"row_heading level0 row6\" >6</th>\n",
              "      <td id=\"T_73b2f_row6_col0\" class=\"data row6 col0\" >The Impact of Artificial Intelligence on Healthcare</td>\n",
              "      <td id=\"T_73b2f_row6_col1\" class=\"data row6 col1\" >Exploring the role of AI in revolutionizing healthcare</td>\n",
              "      <td id=\"T_73b2f_row6_col2\" class=\"data row6 col2\" >artificial-intelligence health</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row7\" class=\"row_heading level0 row7\" >7</th>\n",
              "      <td id=\"T_73b2f_row7_col0\" class=\"data row7 col0\" >The History of Jazz Music</td>\n",
              "      <td id=\"T_73b2f_row7_col1\" class=\"data row7 col1\" >Tracing the origins and evolution of jazz music</td>\n",
              "      <td id=\"T_73b2f_row7_col2\" class=\"data row7 col2\" >history music</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row8\" class=\"row_heading level0 row8\" >8</th>\n",
              "      <td id=\"T_73b2f_row8_col0\" class=\"data row8 col0\" >The Rise of E-commerce in the Digital Age</td>\n",
              "      <td id=\"T_73b2f_row8_col1\" class=\"data row8 col1\" >Examining the growth and impact of online shopping</td>\n",
              "      <td id=\"T_73b2f_row8_col2\" class=\"data row8 col2\" >technology culture</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row9\" class=\"row_heading level0 row9\" >9</th>\n",
              "      <td id=\"T_73b2f_row9_col0\" class=\"data row9 col0\" >The Art of Photography</td>\n",
              "      <td id=\"T_73b2f_row9_col1\" class=\"data row9 col1\" >Exploring the creative and technical aspects of photography</td>\n",
              "      <td id=\"T_73b2f_row9_col2\" class=\"data row9 col2\" >art technology</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row10\" class=\"row_heading level0 row10\" >10</th>\n",
              "      <td id=\"T_73b2f_row10_col0\" class=\"data row10 col0\" >The Psychology of Decision Making</td>\n",
              "      <td id=\"T_73b2f_row10_col1\" class=\"data row10 col1\" >Understanding the cognitive processes behind decision making</td>\n",
              "      <td id=\"T_73b2f_row10_col2\" class=\"data row10 col2\" >psychology</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row11\" class=\"row_heading level0 row11\" >11</th>\n",
              "      <td id=\"T_73b2f_row11_col0\" class=\"data row11 col0\" >The Role of Women in STEM Fields</td>\n",
              "      <td id=\"T_73b2f_row11_col1\" class=\"data row11 col1\" >Highlighting the contributions of women in science, technology, engineering, and mathematics</td>\n",
              "      <td id=\"T_73b2f_row11_col2\" class=\"data row11 col2\" >feminism science technology</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_73b2f_level0_row12\" class=\"row_heading level0 row12\" >12</th>\n",
              "      <td id=\"T_73b2f_row12_col0\" class=\"data row12 col0\" >The Cultural Significance of Tattoos</td>\n",
              "      <td id=\"T_73b2f_row12_col1\" class=\"data row12 col1\" >Exploring the history and symbolism of tattoos in different cultures</td>\n",
              "      <td id=\"T_73b2f_row12_col2\" class=\"data row12 col2\" >culture history</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "                                                  title                                                                                      subtitle                            tags\n",
              "0                                      The Great Gatsby                                                                A novel by F. Scott Fitzgerald                      literature\n",
              "1                              The Theory of Relativity                                                        A scientific theory by Albert Einstein                 science physics\n",
              "2            The Technology and Culture of Ancient Rome                                                    A cross-disciplinary study of ancient Rome      technology culture history\n",
              "3                                Football on Television                               The technology and cultural impact of televising football games       technology culture sports\n",
              "4                        The Philosophy of Taylor Swift                              A philosophical analysis of the music and lyrics of Taylor Swift                philosophy music\n",
              "5                 The Spanish Language in popular music                                  A review of the use of the Spanish language in popular music                  language music\n",
              "6   The Impact of Artificial Intelligence on Healthcare                                        Exploring the role of AI in revolutionizing healthcare  artificial-intelligence health\n",
              "7                             The History of Jazz Music                                               Tracing the origins and evolution of jazz music                   history music\n",
              "8             The Rise of E-commerce in the Digital Age                                            Examining the growth and impact of online shopping              technology culture\n",
              "9                                The Art of Photography                                   Exploring the creative and technical aspects of photography                  art technology\n",
              "10                    The Psychology of Decision Making                                  Understanding the cognitive processes behind decision making                      psychology\n",
              "11                     The Role of Women in STEM Fields  Highlighting the contributions of women in science, technology, engineering, and mathematics     feminism science technology\n",
              "12                 The Cultural Significance of Tattoos                          Exploring the history and symbolism of tattoos in different cultures                 culture history"
            ]
          },
          "execution_count": 12,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "items = [\n",
        "    {\"title\": \"The Great Gatsby\", \"subtitle\": \"A novel by F. Scott Fitzgerald\"},\n",
        "    {\"title\": \"The Theory of Relativity\", \"subtitle\": \"A scientific theory by Albert Einstein\"},\n",
        "    {\"title\": \"The Technology and Culture of Ancient Rome\", \"subtitle\": \"A cross-disciplinary study of ancient Rome\"},\n",
        "    {\"title\": \"Football on Television\", \"subtitle\": \"The technology and cultural impact of televising football games\"},\n",
        "    {\"title\": \"The Philosophy of Taylor Swift\", \"subtitle\": \"A philosophical analysis of the music and lyrics of Taylor Swift\"},\n",
        "    {\"title\": \"The Spanish Language in popular music\", \"subtitle\": \"A review of the use of the Spanish language in popular music\"},\n",
        "    {\"title\": \"The Impact of Artificial Intelligence on Healthcare\", \"subtitle\": \"Exploring the role of AI in revolutionizing healthcare\"},\n",
        "    {\"title\": \"The History of Jazz Music\", \"subtitle\": \"Tracing the origins and evolution of jazz music\"},\n",
        "    {\"title\": \"The Rise of E-commerce in the Digital Age\", \"subtitle\": \"Examining the growth and impact of online shopping\"},\n",
        "    {\"title\": \"The Art of Photography\", \"subtitle\": \"Exploring the creative and technical aspects of photography\"},\n",
        "    {\"title\": \"The Psychology of Decision Making\", \"subtitle\": \"Understanding the cognitive processes behind decision making\"},\n",
        "    {\"title\": \"The Role of Women in STEM Fields\", \"subtitle\": \"Highlighting the contributions of women in science, technology, engineering, and mathematics\"},\n",
        "    {\"title\": \"The Cultural Significance of Tattoos\", \"subtitle\": \"Exploring the history and symbolism of tattoos in different cultures\"},\n",
        "]\n",
        "\n",
        "\n",
        "class Tag(str, Enum):\n",
        "    literature = \"literature\"\n",
        "    science = \"science\"\n",
        "    history = \"history\"\n",
        "    technology = \"technology\"\n",
        "    art = \"art\"\n",
        "    music = \"music\"\n",
        "    sports = \"sports\"\n",
        "    philosophy = \"philosophy\"\n",
        "    language = \"language\"\n",
        "    feminism = \"feminism\"\n",
        "    health = \"health\"\n",
        "    media = \"media\"\n",
        "    physics = \"physics\"\n",
        "    culture = \"culture\"\n",
        "    psychology = \"psychology\"\n",
        "    artificial_intelligence = \"artificial-intelligence\"\n",
        "\n",
        "\n",
        "class TaggableItem(BaseModel):\n",
        "    title: str\n",
        "    subtitle: str\n",
        "    tags: List[Tag]\n",
        "\n",
        "\n",
        "class TaggableItems(BaseModel):\n",
        "    items: List[TaggableItem]\n",
        "\n",
        "\n",
        "tagged_items = llm(\n",
        "    TaggableItems,\n",
        "    f\"Tag the following items with the appropriate tags. Options: {json.dumps([tag.value for tag in Tag])}\",\n",
        "    json.dumps(items),\n",
        ")\n",
        "\n",
        "pd.DataFrame([\n",
        "  {\"title\": item.title, \"subtitle\": item.subtitle, \"tags\": ' '.join(item.tags)}\n",
        "  for item in tagged_items.items\n",
        "])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Clustering\n",
        "\n",
        "Now that we have classified out items and assigned a tag to each of them, we might want to cluster them together, based on their content and the tags assigned. One of the advantages of using an LLM to complete this task (rather than a predictive model), is that the LLM can also explain the choices it made, for example by giving each cluster a title."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 13,
      "metadata": {},
      "outputs": [
        {
          "data": {
            "text/html": [
              "<style type=\"text/css\">\n",
              "#T_72e69_row0_col0, #T_72e69_row0_col1, #T_72e69_row0_col2, #T_72e69_row0_col3, #T_72e69_row1_col0, #T_72e69_row1_col1, #T_72e69_row1_col2, #T_72e69_row1_col3, #T_72e69_row2_col0, #T_72e69_row2_col1, #T_72e69_row2_col2, #T_72e69_row2_col3, #T_72e69_row3_col0, #T_72e69_row3_col1, #T_72e69_row3_col2, #T_72e69_row3_col3, #T_72e69_row4_col0, #T_72e69_row4_col1, #T_72e69_row4_col2, #T_72e69_row4_col3, #T_72e69_row5_col0, #T_72e69_row5_col1, #T_72e69_row5_col2, #T_72e69_row5_col3, #T_72e69_row6_col0, #T_72e69_row6_col1, #T_72e69_row6_col2, #T_72e69_row6_col3, #T_72e69_row7_col0, #T_72e69_row7_col1, #T_72e69_row7_col2, #T_72e69_row7_col3, #T_72e69_row8_col0, #T_72e69_row8_col1, #T_72e69_row8_col2, #T_72e69_row8_col3, #T_72e69_row9_col0, #T_72e69_row9_col1, #T_72e69_row9_col2, #T_72e69_row9_col3, #T_72e69_row10_col0, #T_72e69_row10_col1, #T_72e69_row10_col2, #T_72e69_row10_col3, #T_72e69_row11_col0, #T_72e69_row11_col1, #T_72e69_row11_col2, #T_72e69_row11_col3, #T_72e69_row12_col0, #T_72e69_row12_col1, #T_72e69_row12_col2, #T_72e69_row12_col3 {\n",
              "  text-align: left;\n",
              "}\n",
              "</style>\n",
              "<table id=\"T_72e69\">\n",
              "  <thead>\n",
              "    <tr>\n",
              "      <th class=\"blank level0\" >&nbsp;</th>\n",
              "      <th id=\"T_72e69_level0_col0\" class=\"col_heading level0 col0\" >cluster</th>\n",
              "      <th id=\"T_72e69_level0_col1\" class=\"col_heading level0 col1\" >title</th>\n",
              "      <th id=\"T_72e69_level0_col2\" class=\"col_heading level0 col2\" >subtitle</th>\n",
              "      <th id=\"T_72e69_level0_col3\" class=\"col_heading level0 col3\" >tags</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
              "      <td id=\"T_72e69_row0_col0\" class=\"data row0 col0\" >Literature and Philosophy</td>\n",
              "      <td id=\"T_72e69_row0_col1\" class=\"data row0 col1\" >The Great Gatsby</td>\n",
              "      <td id=\"T_72e69_row0_col2\" class=\"data row0 col2\" >A novel by F. Scott Fitzgerald</td>\n",
              "      <td id=\"T_72e69_row0_col3\" class=\"data row0 col3\" >literature</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
              "      <td id=\"T_72e69_row1_col0\" class=\"data row1 col0\" >Literature and Philosophy</td>\n",
              "      <td id=\"T_72e69_row1_col1\" class=\"data row1 col1\" >The Philosophy of Taylor Swift</td>\n",
              "      <td id=\"T_72e69_row1_col2\" class=\"data row1 col2\" >A philosophical analysis of the music and lyrics of Taylor Swift</td>\n",
              "      <td id=\"T_72e69_row1_col3\" class=\"data row1 col3\" >philosophy music</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
              "      <td id=\"T_72e69_row2_col0\" class=\"data row2 col0\" >Science and Psychology</td>\n",
              "      <td id=\"T_72e69_row2_col1\" class=\"data row2 col1\" >The Theory of Relativity</td>\n",
              "      <td id=\"T_72e69_row2_col2\" class=\"data row2 col2\" >A scientific theory by Albert Einstein</td>\n",
              "      <td id=\"T_72e69_row2_col3\" class=\"data row2 col3\" >science physics</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
              "      <td id=\"T_72e69_row3_col0\" class=\"data row3 col0\" >Science and Psychology</td>\n",
              "      <td id=\"T_72e69_row3_col1\" class=\"data row3 col1\" >The Psychology of Decision Making</td>\n",
              "      <td id=\"T_72e69_row3_col2\" class=\"data row3 col2\" >Understanding the cognitive processes behind decision making</td>\n",
              "      <td id=\"T_72e69_row3_col3\" class=\"data row3 col3\" >psychology</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
              "      <td id=\"T_72e69_row4_col0\" class=\"data row4 col0\" >Technology and Culture</td>\n",
              "      <td id=\"T_72e69_row4_col1\" class=\"data row4 col1\" >The Technology and Culture of Ancient Rome</td>\n",
              "      <td id=\"T_72e69_row4_col2\" class=\"data row4 col2\" >A cross-disciplinary study of ancient Rome</td>\n",
              "      <td id=\"T_72e69_row4_col3\" class=\"data row4 col3\" >technology culture history</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row5\" class=\"row_heading level0 row5\" >5</th>\n",
              "      <td id=\"T_72e69_row5_col0\" class=\"data row5 col0\" >Technology and Culture</td>\n",
              "      <td id=\"T_72e69_row5_col1\" class=\"data row5 col1\" >Football on Television</td>\n",
              "      <td id=\"T_72e69_row5_col2\" class=\"data row5 col2\" >The technology and cultural impact of televising football games</td>\n",
              "      <td id=\"T_72e69_row5_col3\" class=\"data row5 col3\" >technology culture sports</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row6\" class=\"row_heading level0 row6\" >6</th>\n",
              "      <td id=\"T_72e69_row6_col0\" class=\"data row6 col0\" >Technology and Culture</td>\n",
              "      <td id=\"T_72e69_row6_col1\" class=\"data row6 col1\" >The Rise of E-commerce in the Digital Age</td>\n",
              "      <td id=\"T_72e69_row6_col2\" class=\"data row6 col2\" >Examining the growth and impact of online shopping</td>\n",
              "      <td id=\"T_72e69_row6_col3\" class=\"data row6 col3\" >technology culture</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row7\" class=\"row_heading level0 row7\" >7</th>\n",
              "      <td id=\"T_72e69_row7_col0\" class=\"data row7 col0\" >Technology and Culture</td>\n",
              "      <td id=\"T_72e69_row7_col1\" class=\"data row7 col1\" >The Art of Photography</td>\n",
              "      <td id=\"T_72e69_row7_col2\" class=\"data row7 col2\" >Exploring the creative and technical aspects of photography</td>\n",
              "      <td id=\"T_72e69_row7_col3\" class=\"data row7 col3\" >art technology</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row8\" class=\"row_heading level0 row8\" >8</th>\n",
              "      <td id=\"T_72e69_row8_col0\" class=\"data row8 col0\" >Music and Language</td>\n",
              "      <td id=\"T_72e69_row8_col1\" class=\"data row8 col1\" >The Spanish Language in popular music</td>\n",
              "      <td id=\"T_72e69_row8_col2\" class=\"data row8 col2\" >A review of the use of the Spanish language in popular music</td>\n",
              "      <td id=\"T_72e69_row8_col3\" class=\"data row8 col3\" >language music</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row9\" class=\"row_heading level0 row9\" >9</th>\n",
              "      <td id=\"T_72e69_row9_col0\" class=\"data row9 col0\" >Music and Language</td>\n",
              "      <td id=\"T_72e69_row9_col1\" class=\"data row9 col1\" >The History of Jazz Music</td>\n",
              "      <td id=\"T_72e69_row9_col2\" class=\"data row9 col2\" >Tracing the origins and evolution of jazz music</td>\n",
              "      <td id=\"T_72e69_row9_col3\" class=\"data row9 col3\" >history music</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row10\" class=\"row_heading level0 row10\" >10</th>\n",
              "      <td id=\"T_72e69_row10_col0\" class=\"data row10 col0\" >Health and Feminism</td>\n",
              "      <td id=\"T_72e69_row10_col1\" class=\"data row10 col1\" >The Impact of Artificial Intelligence on Healthcare</td>\n",
              "      <td id=\"T_72e69_row10_col2\" class=\"data row10 col2\" >Exploring the role of AI in revolutionizing healthcare</td>\n",
              "      <td id=\"T_72e69_row10_col3\" class=\"data row10 col3\" >artificial-intelligence health</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row11\" class=\"row_heading level0 row11\" >11</th>\n",
              "      <td id=\"T_72e69_row11_col0\" class=\"data row11 col0\" >Health and Feminism</td>\n",
              "      <td id=\"T_72e69_row11_col1\" class=\"data row11 col1\" >The Role of Women in STEM Fields</td>\n",
              "      <td id=\"T_72e69_row11_col2\" class=\"data row11 col2\" >Highlighting the contributions of women in science, technology, engineering, and mathematics</td>\n",
              "      <td id=\"T_72e69_row11_col3\" class=\"data row11 col3\" >feminism science technology</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th id=\"T_72e69_level0_row12\" class=\"row_heading level0 row12\" >12</th>\n",
              "      <td id=\"T_72e69_row12_col0\" class=\"data row12 col0\" >Health and Feminism</td>\n",
              "      <td id=\"T_72e69_row12_col1\" class=\"data row12 col1\" >The Cultural Significance of Tattoos</td>\n",
              "      <td id=\"T_72e69_row12_col2\" class=\"data row12 col2\" >Exploring the history and symbolism of tattoos in different cultures</td>\n",
              "      <td id=\"T_72e69_row12_col3\" class=\"data row12 col3\" >culture history</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n"
            ],
            "text/plain": [
              "                      cluster                                                title                                                                                      subtitle                            tags\n",
              "0   Literature and Philosophy                                     The Great Gatsby                                                                A novel by F. Scott Fitzgerald                      literature\n",
              "1   Literature and Philosophy                       The Philosophy of Taylor Swift                              A philosophical analysis of the music and lyrics of Taylor Swift                philosophy music\n",
              "2      Science and Psychology                             The Theory of Relativity                                                        A scientific theory by Albert Einstein                 science physics\n",
              "3      Science and Psychology                    The Psychology of Decision Making                                  Understanding the cognitive processes behind decision making                      psychology\n",
              "4      Technology and Culture           The Technology and Culture of Ancient Rome                                                    A cross-disciplinary study of ancient Rome      technology culture history\n",
              "5      Technology and Culture                               Football on Television                               The technology and cultural impact of televising football games       technology culture sports\n",
              "6      Technology and Culture            The Rise of E-commerce in the Digital Age                                            Examining the growth and impact of online shopping              technology culture\n",
              "7      Technology and Culture                               The Art of Photography                                   Exploring the creative and technical aspects of photography                  art technology\n",
              "8          Music and Language                The Spanish Language in popular music                                  A review of the use of the Spanish language in popular music                  language music\n",
              "9          Music and Language                            The History of Jazz Music                                               Tracing the origins and evolution of jazz music                   history music\n",
              "10        Health and Feminism  The Impact of Artificial Intelligence on Healthcare                                        Exploring the role of AI in revolutionizing healthcare  artificial-intelligence health\n",
              "11        Health and Feminism                     The Role of Women in STEM Fields  Highlighting the contributions of women in science, technology, engineering, and mathematics     feminism science technology\n",
              "12        Health and Feminism                 The Cultural Significance of Tattoos                          Exploring the history and symbolism of tattoos in different cultures                 culture history"
            ]
          },
          "execution_count": 13,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "num_clusters = 5\n",
        "\n",
        "class TaggableItemsCluster(BaseModel):\n",
        "    title: str = Field(..., description=\"The title of the cluster.\")\n",
        "    items: List[str] = Field(..., description=\"The titles of the items in the cluster.\")\n",
        "\n",
        "class TaggableItemsClusters(BaseModel):\n",
        "    clusters: List[TaggableItemsCluster]\n",
        "\n",
        "tagged_items_clusters = llm(\n",
        "    TaggableItemsClusters,\n",
        "    f\"Cluster the following items based on their tags and content. Create exactly {num_clusters} clusters.\",\n",
        "    json.dumps(tagged_items.dict()[\"items\"]),\n",
        ")\n",
        "\n",
        "clusters = []\n",
        "for cluster in tagged_items_clusters.clusters:\n",
        "  for item_title in cluster.items:\n",
        "    item = next(item for item in tagged_items.items if item.title == item_title)\n",
        "    clusters.append({\n",
        "       \"cluster\": cluster.title,\n",
        "       \"title\": item.title,\n",
        "       \"subtitle\": item.subtitle,\n",
        "       \"tags\": ' '.join(item.tags)\n",
        "    })\n",
        "pd.DataFrame(clusters)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "💡 What other use-cases can you think of where simple decision-making can be handed off to an LLM? Can you think of examples where you'd be OK with letting the LLM make some decisions without human inspection? How will you know if the accuracy is good enough?"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "pIbTVEz9kQry"
      },
      "source": [
        "## Planning and Tool-Use\n",
        "\n",
        "Complex systems and behaviours often need to plan multiple steps ahead and interact with the \"world\". LLMs can often do that quite well. Let's look at a couple of examples.\n",
        "\n",
        "If our LLM knows of a distinct set of actions it can take, we can get it to plan which actions to perform and in what order, based on the relevant situation."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 14,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Activity: Waking up and going to work\n",
            "1. Wake up\n",
            "2. Turn off the alarm\n",
            "3. Stretch\n",
            "4. Get out of bed\n",
            "5. Use the bathroom\n",
            "6. Wash face\n",
            "7. Brush teeth\n",
            "8. Get dressed\n",
            "9. Apply deodorant\n",
            "10. Comb or brush hair\n",
            "11. Prepare breakfast\n",
            "12. Eat breakfast\n",
            "13. Make coffee or tea\n",
            "14. Check phone for messages or emails\n",
            "15. Pack lunch\n",
            "16. Gather work materials\n",
            "17. Put on shoes\n",
            "18. Grab keys\n",
            "19. Lock the door\n",
            "20. Leave the house\n",
            "\n",
            "Activity: Winding down and going to sleep\n",
            "1. Finish dinner\n",
            "2. Clean up dinner dishes\n",
            "3. Watch TV or read a book\n",
            "4. Brush teeth in the evening\n",
            "5. Use the bathroom in the evening\n",
            "6. Wash face in the evening\n",
            "7. Change into pyjamas\n",
            "8. Set alarm for the next day\n",
            "9. Turn off electronic devices\n",
            "10. Get into bed\n",
            "11. Adjust pillows and blankets\n",
            "12. Close eyes and try to sleep\n",
            "\n",
            "Activity: Going to see a movie with friends\n",
            "1. Decide on a movie to watch\n",
            "2. Check movie times online\n",
            "3. Decide on meeting place and time\n",
            "4. Call a taxi or arrange transportation\n",
            "5. Meet friends at designated place\n",
            "6. Walk or drive to the movie theatre\n",
            "7. Purchase tickets at the theatre\n",
            "8. Buy snacks at the concession stand\n",
            "9. Find the correct theatre screen\n",
            "10. Find seats\n",
            "11. Watch the movie\n",
            "12. Discuss the movie with friends\n",
            "13. Say goodbye to friends\n",
            "14. Return home\n",
            "\n"
          ]
        }
      ],
      "source": [
        "class Action(str, Enum):\n",
        "  WAKE_UP = \"Wake up\"\n",
        "  TURN_OFF_ALARM = \"Turn off the alarm\"\n",
        "  STRETCH = \"Stretch\"\n",
        "  GET_OUT_OF_BED = \"Get out of bed\"\n",
        "  USE_BATHROOM = \"Use the bathroom\"\n",
        "  CHECK_FOR_MOVIE_SNACKS = \"Check for movie snacks\"\n",
        "  WASH_FACE_EVENING = \"Wash face in the evening\"\n",
        "  CHANGE_INTO_PYJAMAS = \"Change into pyjamas\"\n",
        "  SET_ALARM_FOR_NEXT_DAY = \"Set alarm for the next day\"\n",
        "  CHECK_PHONE_FOR_MESSAGES = \"Check phone for messages\"\n",
        "  TURN_OFF_LIGHTS = \"Turn off lights\"\n",
        "  WALK_OR_DRIVE_TO_MOVIE_THEATRE = \"Walk or drive to the movie theatre\"\n",
        "  USE_BATHROOM_EVENING = \"Use the bathroom in the evening\"\n",
        "  DRY_OFF_WITH_TOWEL = \"Dry off with a towel\"\n",
        "  BRUSH_TEETH = \"Brush teeth\"\n",
        "  WASH_FACE = \"Wash face\"\n",
        "  SHOWER = \"Take a shower\"\n",
        "  GRAB_WALLET_PURSE = \"Grab wallet or purse\"\n",
        "  MAKE_SURE_PHONE_IS_CHARGED = \"Make sure phone is charged\"\n",
        "  CALL_A_TAXI_ARRANGE_TRANSPORTATION = \"Call a taxi or arrange transportation\"\n",
        "  MEET_FRIENDS_AT_DESIGNATED_PLACE = \"Meet friends at designated place\"\n",
        "  LEAVE_THE_HOUSE = \"Leave the house\"\n",
        "  PLAN_TO_BUY_AT_THEATRE = \"Plan to buy tickets at the theatre\"\n",
        "  DECIDE_ON_MEETING_PLACE_AND_TIME = \"Decide on meeting place and time\"\n",
        "  GET_DRESSED = \"Get dressed\"\n",
        "  APPLY_DEODORANT = \"Apply deodorant\"\n",
        "  COMB_BRUSH_HAIR = \"Comb or brush hair\"\n",
        "  STYLE_HAIR = \"Style hair\"\n",
        "  SHAVE = \"Shave\"\n",
        "  PUT_ON_CLOTHES = \"Put on clothes\"\n",
        "  APPLY_MAKEUP = \"Apply makeup\"\n",
        "  PREPARE_BREAKFAST = \"Prepare breakfast\"\n",
        "  EAT_BREAKFAST = \"Eat breakfast\"\n",
        "  MAKE_COFFEE_TEA = \"Make coffee or tea\"\n",
        "  CHECK_PHONE_FOR_MESSAGES_EMAILS = \"Check phone for messages or emails\"\n",
        "  PACK_LUNCH = \"Pack lunch\"\n",
        "  GATHER_WORK_MATERIALS = \"Gather work materials\"\n",
        "  PUT_ON_SHOES = \"Put on shoes\"\n",
        "  GRAB_KEYS = \"Grab keys\"\n",
        "  PURCHASE_TICKETS_AT_THEATRE = \"Purchase tickets at the theatre\"\n",
        "  LOCK_THE_DOOR = \"Lock the door\"\n",
        "  FINISH_DINNER = \"Finish dinner\"\n",
        "  CLEAN_UP_DINNER_DISHES = \"Clean up dinner dishes\"\n",
        "  WATCH_TV_READ_BOOK = \"Watch TV or read a book\"\n",
        "  BRUSH_TEETH_EVENING = \"Brush teeth in the evening\"\n",
        "  GET_INTO_BED = \"Get into bed\"\n",
        "  MEDITATE_RELAX = \"Meditate or relax\"\n",
        "  WRITE_IN_JOURNAL = \"Write in journal\"\n",
        "  LISTEN_TO_CALMING_MUSIC = \"Listen to calming music\"\n",
        "  TURN_OFF_ELECTRONIC_DEVICES = \"Turn off electronic devices\"\n",
        "  ADJUST_PILLOWS_AND_BLANKETS = \"Adjust pillows and blankets\"\n",
        "  READ_BOOK = \"Read a book\"\n",
        "  CLOSE_EYES_TRY_TO_SLEEP = \"Close eyes and try to sleep\"\n",
        "  DECIDE_ON_MOVIE_TO_WATCH = \"Decide on a movie to watch\"\n",
        "  CHECK_MOVIE_TIMES_ONLINE = \"Check movie times online\"\n",
        "  PURCHASE_TICKETS_ONLINE = \"Purchase tickets online\"\n",
        "  BUY_SNACKS_AT_CONCESSION_STAND = \"Buy snacks at the concession stand\"\n",
        "  FIND_CORRECT_THEATRE_SCREEN = \"Find the correct theatre screen\"\n",
        "  FIND_SEATS = \"Find seats\"\n",
        "  WATCH_THE_MOVIE = \"Watch the movie\"\n",
        "  DISCUSS_MOVIE_WITH_FRIENDS = \"Discuss the movie with friends\"\n",
        "  SAY_GOODBYE_TO_FRIENDS = \"Say goodbye to friends\"\n",
        "  RETURN_HOME = \"Return home\"\n",
        "\n",
        "activities = [\n",
        "  \"Waking up and going to work\",\n",
        "  \"Winding down and going to sleep\",\n",
        "  \"Going to see a movie with friends\",\n",
        "]\n",
        "\n",
        "class ActionSequence(BaseModel):\n",
        "  actions: List[Action]\n",
        "\n",
        "for activity in activities:\n",
        "  action_sequence = llm(\n",
        "      ActionSequence,\n",
        "      \"Generate a sequence of actions for the user-provided activity.\",\n",
        "      activity,\n",
        "  )\n",
        "  print(f\"Activity: {activity}\")\n",
        "  for index, action in enumerate(action_sequence.actions, start=1):\n",
        "    print(f\"{index}. {action.value}\")\n",
        "  print()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "For our final example, let's get our LLM act as a game-playing engine. The game is simple, tic-tac-toe (that's a simple example, but the best LLMs have been shown capable of playing much more complex games). We can get a lot of behaviour with very little programming, just by asking the LLM and restricting the input and output."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 15,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "X| | \n",
            "-----\n",
            " | | \n",
            "-----\n",
            " | | \n",
            "\n",
            "X|O| \n",
            "-----\n",
            " | | \n",
            "-----\n",
            " | | \n",
            "\n",
            "X|O|X\n",
            "-----\n",
            " | | \n",
            "-----\n",
            " | | \n",
            "\n",
            "X|O|X\n",
            "-----\n",
            "O| | \n",
            "-----\n",
            " | | \n",
            "\n",
            "X|O|X\n",
            "-----\n",
            "O|X| \n",
            "-----\n",
            " | | \n",
            "\n",
            "X|O|X\n",
            "-----\n",
            "O|X|O\n",
            "-----\n",
            " | | \n",
            "\n",
            "X|O|X\n",
            "-----\n",
            "O|X|O\n",
            "-----\n",
            " | |X\n",
            "\n",
            "Game Over! Winner: X\n"
          ]
        }
      ],
      "source": [
        "class TicTacToeMove(BaseModel):\n",
        "  row: int\n",
        "  col: int\n",
        "\n",
        "class TicTacToeStrategy(str, Enum):\n",
        "  optimal = \"Optimal. Always choose the best move for winning the game or preventing your opponent from winning.\"\n",
        "  random = \"Random. Choose your next move at random.\"\n",
        "  next_free = \"Next Freee. Always choose the next free spot counting from the top-left.\"\n",
        "\n",
        "class TicTacToeWinner(str, Enum):\n",
        "  X = \"X\"\n",
        "  O = \"O\"\n",
        "  Tie = \"Tie\"\n",
        "  Ongoing = \"Ongoing\"\n",
        "\n",
        "class TicTacToeStatus(BaseModel):\n",
        "  winner: TicTacToeWinner\n",
        "\n",
        "class TicTacToeBoard:\n",
        "  def __init__(self):\n",
        "    self.board = [[' ' for _ in range(3)] for _ in range(3)]\n",
        "  \n",
        "  def dumps_board(self):\n",
        "    return '\\n-----\\n'.join(['|'.join(row) for row in self.board]) + '\\n'\n",
        "\n",
        "  def print_board(self):\n",
        "    print(self.dumps_board())\n",
        "  \n",
        "  def make_move(self, role, move: TicTacToeMove):\n",
        "    self.board[move.row][move.col] = role\n",
        "  \n",
        "  def check_status(self) -> TicTacToeWinner:\n",
        "    return llm(\n",
        "        TicTacToeStatus,\n",
        "        (\"Check the status and whether there is a winner in this game of Tic-Tac-Toe. \" +\n",
        "         \"The options are: X - X won, O - O won, \" +\n",
        "         \"Tie - there are no more possible moves and it's a tie, \" +\n",
        "         \"Ongoing - there are still empty slots on the board.\"),\n",
        "        self.dumps_board(),\n",
        "        max_tokens=50,\n",
        "    ).winner\n",
        "\n",
        "class TicTacToePlayer:\n",
        "  def __init__(self, role, strategy):\n",
        "    self.role = role\n",
        "    self.strategy = strategy\n",
        "\n",
        "  def turn(self, board):\n",
        "    move = llm(\n",
        "        TicTacToeMove,\n",
        "        (\"You are a Tic-Tac-Toe Player. \" +\n",
        "         f\"Your role is: {self.role}. It's your turn. \" +\n",
        "         f\"Your strategy is: {self.strategy}. \"\n",
        "         \"Look at the board and announce your move.\"),\n",
        "        board.dumps_board(),\n",
        "        max_tokens = 50,\n",
        "    )\n",
        "    board.make_move(self.role, move)\n",
        "    board.print_board()\n",
        "\n",
        "board = TicTacToeBoard()\n",
        "player_x = TicTacToePlayer('X', TicTacToeStrategy.optimal)\n",
        "player_o = TicTacToePlayer('O', TicTacToeStrategy.next_free)\n",
        "\n",
        "next_player = player_x\n",
        "while board.check_status() == TicTacToeWinner.Ongoing:\n",
        "  next_player.turn(board)\n",
        "  next_player = player_x if next_player == player_o else player_o\n",
        "\n",
        "print(f\"Game Over! Winner: {board.check_status()}\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "💡 Can you think of more demanding tasks you could automate by connecting an LLM to the \"real world\"? How can you structure the interface between the LLM and external \"tools\"? How will you validate that the LLM's decisions are optimal, or at least ensure that you can recover from sub-optimal decisions?"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Conclusion - Your LLM is a VM\n",
        "\n",
        "In this tutorial we explored simple ways of using an LLM as software device. One that gets programmed with instructions in human language, but also understands and respects software interfaces. These examples may seem trivial, but the best LLMs, like GPT-4o, can handle a lot more complexity, and perform with quite a lot of \"intelligence\". Now that you're familiar with this way of thinking about LLMs, and with some techniques for interfacing with an LLM from your software, what will you build?"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "provenance": []
    },
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.10.13"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}