OpenAI request NotFoundError response adds additional context to in…

…dicate potential lack of funds (#784)

When an OpenAI token is used that either does not have billing enabled
or does not have enough credits to call the default model, their API
will raise an exceptionof type `NotFoundError`.

When this happens, we will catch the errors of this type and modify the
message to include additional context that indicates that the issue may
be related to insufficient credits. We will then re-raise an
`APUException`.

Fixes issue #709

---------

Co-authored-by: Luke Lalor <lukehlalor@gmail.com>

sdk/eidolon_ai_sdk/apu/llm/open_ai_connection_handler.py

@@ -2,12 +2,13 @@
 from typing import Optional, cast, List
 
 from azure.identity import get_bearer_token_provider, EnvironmentCredential
-from openai import AsyncOpenAI, AsyncStream
+from openai import AsyncOpenAI, AsyncStream, NotFoundError
 from openai.lib.azure import AsyncAzureOpenAI
 from openai.types import ImagesResponse
 from openai.types.chat import ChatCompletionChunk, ChatCompletion
 from pydantic import BaseModel, Field
 
+from eidolon_ai_sdk.apu.apu import APUException
 from eidolon_ai_sdk.system.reference_model import Specable, Reference
 from eidolon_ai_sdk.util.replay import replayable
 
@@ -22,7 +23,7 @@ def makeClient(self) -> AsyncOpenAI:
 
     async def completion(self, **kwargs) -> ChatCompletion | AsyncStream[ChatCompletionChunk]:
         return await replayable(
-            fn=lambda **_kwargs: self.makeClient().chat.completions.create(**_kwargs),
+            fn=lambda **_kwargs: self._make_request(**_kwargs),
             parser=_replay_parser,
             name_override="openai_completion",
         )(**kwargs)
@@ -31,6 +32,25 @@ async def generate_image(self, **kwargs) -> ImagesResponse:
         # todo, image generation should be repayable, but needs custom parser
         return await self.makeClient().images.generate(**kwargs)
 
+    async def _make_request(self, **kwargs):
+        try:
+            return await self.makeClient().chat.completions.create(**kwargs)
+        except NotFoundError as e:
+            # The user likely does not have access to the requested named model or the model name is invalid.
+            # A typical response looks like this:
+            # {
+            #   "error": {
+            #     "message": "The model `gpt-4-turbo` does not exist or you do not have access to it.",
+            #     "type": "invalid_request_error",
+            #     "code": "model_not_found"
+            #   }
+            # }
+            message = (
+                f"{e.message}\n" 
+                "OpenAI accounts that do not have sufficient credits may not have access to some models."
+            )
+            raise APUException(description=message)
+
 
 def get_default_token_provider():
     if os.environ.get("AZURE_CLIENT_ID") and os.environ.get("AZURE_CLIENT_SECRET") and os.environ.get("AZURE_TENANT_ID"):

sdk/tests/apu/llm/test_open_ai_connection_handler.py

@@ -0,0 +1,42 @@
+import pytest
+from pytest import fail
+from httpx import Response, Request
+from openai import AsyncOpenAI, NotFoundError
+
+from eidolon_ai_sdk.apu.apu import APUException
+from eidolon_ai_sdk.apu.llm.open_ai_connection_handler import OpenAIConnectionHandler
+
+
+class TestOpenAIConnectionHandler(OpenAIConnectionHandler):
+    def makeClient(self) -> AsyncOpenAI:
+        raise NotFoundError(
+            self.connection_response_message(),
+            response=Response(
+                404,
+                request=Request("POST", "https://example.com")
+            ),
+            body=None,
+        )
+
+    def connection_response_message(self):
+        return "The model `gpt-4-turbo` does not exist or you do not have access to it."
+
+
+@pytest.fixture(scope="module")
+async def connection():
+    # noinspection PyTypeChecker
+    return TestOpenAIConnectionHandler(spec=None)
+
+
+async def test_not_found_response(connection, **kwargs):
+    """
+    Asserts that when a NotFoundException is thrown by the OpenAI client then an appropriate message is returned
+    to the user as part of an APUException
+    """
+    try:
+        await connection.completion(**kwargs)
+        fail("Expected to raise APUException")
+    except APUException as e:
+        message = str(e)
+        assert connection.connection_response_message() in message
+        assert "OpenAI accounts that do not have sufficient credits may not have access to some models." in message