Skip to content

Latest commit

 

History

History
60 lines (60 loc) · 2.64 KB

TODO.org

File metadata and controls

60 lines (60 loc) · 2.64 KB

Really TODO

sign-up page

Dockerfile (#sadface)

use client context `now` for memory sql dates

Implement websockets for chat

documentation (docstring & readmes)

better intro flow

/help command

/context command

export / download data functionality (for me and the GDPR!)

calendar page (for memories)

#keyword search

semantic search memories

do not use JSON for Esther responses (no need really with gbnf constraints)

  • It confuses the llm as it pushes the logits towards json garbage, rather than natural language

test infrastructure and tests => generate training / test data for fine-tuning

embeddings for RAG (memory retrieval for context)

  • onnx bert?
  • bert.cpp / jna?
  • using llama.cpp embeddings directly?
  • Add memories during converse into associative mem
  • Retrieve similar memories (if any) => guids (associate)
  • For prompt: frecency keywords => memories by keyword => most_similar => summarize top-k => context: {kw: summary}
  • For converse: most_similar => summarize top-k => context
  • Inject similar memories in user request context
  • https://github.com/nmslib/hnswlib (via JNA clong clang?)
  • https://github.com/vioshyvo/mrpt (no persistence)
  • get it to work simple example
  • get it to work persistence, how to deal with scoping to user?!
  • get it to work with embeddings and real data (find or generate some dataset)
  • summarize using simple lexrank on imaginations?
  • how to generate the proper amount of context? (summarizing???)

UI/UX things

  • [X] Discoverability of input bar — Maybe show imagination?
  • [X] Nicer / intelligent placeholder?
  • [ ] Button for archive or other commands?
  • [ ] Dark mode
  • [ ] Logo + favicon
  • [X] Loading indicator login
  • [ ] Accessibility and ARIA
  • [X] Disable text input while in modal confirm

llama.cpp parallel decode?

share a page (static gen)

security and hosting

deploy and build shizzle (Github actions?)

app? legal, ethics, billing, subscriptions, etc.

Ideas for later

finetune model

link parsing in that it downloads the page and adds it as context

image upload (img-to-text?)

offline desktop app?

/imagine command (no solutions for this ATM)

  • StableDiffusion locally?
  • maybe StableDiffusion v0.9 (new version)
  • integrate image describe into memory
  • https://github.com/deep-floyd/IF maybe?
  • store description and image in memory contents
  • use image describe as context for converse

Simularca for data generation / simulations