TODOS

[louis030195]

• [ ] general repo
  • [x] ci cd
  • [ ] good readme
  • [ ] instructions contribute etc
  • [ ] templates
  • [ ] consider generating models like openai based on the yaml thing
  • [ ] tests that run on openai api then on open source assistants and compare results
  • [x] properly segmented by users
  • [ ] split assistants-core code in smaller chunks
  • [x] docker stuff
  • [x] software 3.0 stuff
  • [x] basically openai assistants takes automatic decisions on tools to use - gotta implement something smooth that well replicate (need some more data on the experience if anyone is big user of their api?)
  • [ ] all the entrypoints
  • [x] POST /assistants - Create a new assistant.
  • [x] GET /assistants/{assistant_id} - Retrieve a specific assistant by its ID.
  • [x] GET /assistants - List all assistants.
  • [x] POST /assistants/{assistant_id} - Update a specific assistant.
  • [x] DELETE /assistants/{assistant_id} - Delete a specific assistant.
  • [x] POST /threads - Create a new thread.
  • [x] GET /threads/{thread_id} - Retrieve a specific thread by its ID.
  • [x] GET /threads - List all threads.
  • [x] POST /threads/{thread_id} - Update a specific thread.
  • [x] DELETE /threads/{thread_id} - Delete a specific thread.
  • [x] POST /threads/{thread_id}/messages - Add a new message to a thread.
  • [x] GET /threads/{thread_id}/messages/{message_id} - Retrieve a specific message by its ID.
  • [x] GET /threads/{thread_id}/messages - List all messages in a thread.
  • [x] DELETE /threads/{thread_id}/messages/{message_id} - Delete a specific message.
  • [x] POST /threads/{thread_id}/runs - Create a run.
  • [x] GET /threads/{thread_id}/runs/{run_id} - Retrieve a run.
  • [x] POST /threads/{thread_id}/runs/{run_id} - Modify a run.
  • [x] GET /threads/{thread_id}/runs - List runs.
  • [x] POST /threads/{thread_id}/runs/{run_id}/submit_tool_outputs - Submit tool outputs to run.
  • [ ] POST /threads/{thread_id}/runs/{run_id}/cancel - Cancel a run.
  • [ ] POST /threads/runs - Create thread and run.
  • [ ] GET /threads/{thread_id}/runs/{run_id}/steps/{step_id} - Retrieve run step.
  • [ ] GET /threads/{thread_id}/runs/{run_id}/steps - List run steps.
  • [x] POST /files - Upload a file.
  • [ ] GET /files/{file_id} - Fetch a file's details.
  • [ ] POST /files/{file_id} - Update a file's details.
  • [ ] DELETE /files/{file_id} - Delete a file.
  • [ ] GET /files - List all files.
  • [ ] GET /files/{file_id}/download - Download a file.
  • [ ] 10/10 error handling (atm 4/10)
  • [ ] run error
  • [ ] general error handling
    • [ ] shouldnt crash the api when no rows are returned :) (unit test!!)
  • [ ] all the properties in models used properly like openai
• [x] end to end overview
  • [x] logic
  • [x] api
  • [x] models
• [ ] retrieval
  • [ ] examples
  • [x] curl anthropic hello world
  • [x] curl mistral hello world
  • [x] js retrieval anthropic hello world
  • [ ] nextjs hello world
  • [ ] annotations
  • [ ] vector
  • [ ] in pg?
  • [ ] in redis?
  • [ ] external vector db? (fuck additional infra :D)
  • [ ] full text
  • [ ] sql generation / execution (out of project scope? could be an example)
• [x] Files
  • [x] pdfs & co
  • [ ] images, jpg ...
  • [ ] video
  • [x] txt
• [ ] code interpreter
  • [ ] suggest in a new dir like "assistants-code-interpreter" because i assume there will be a bunch of specific deps to execute code etc, scalable thing - maybe some copy paste from there
• [x] function calling
  • [x] anthropic
  • [x] openai
  • [x] open source llm
  • [x] basic tests
  • [ ] bunch of tests for all edge cases
  • [ ] examples
  • [x] end-to-end thru curl anthropic
  • [x] end-to-end intel neural-chat nodejs
  • [ ] end-to-end curl mistral
  • [ ] end-to-end curl mistral retrieval + function calling
• [ ] LLMs
  • [x] OpenAI (useful for open source LLMs deployed with OpenAI like API)
  • [x] Anthropic
  • [x] open source
  • [x] document how to run LLMs (vLLM, etc.)
  • [x] fastchat (mac/cpu)
  • [ ] perplexity
  • [ ] vllm (gpu)
  • [ ] one click cloud
  • [ ] k8s gpu
  • [ ] modal
  • [ ] replicate
  • [ ] skypilot
• [ ] infra
  • [ ] one-click to deploy everything somewhere in ur infra in less than 10 seconds
  • [ ] enterprise support (k8s)
  • [x] v0
  • [ ] readme example k8s in < 1 min
  • [ ] on the edge (not arm!) without internet access
• [ ] performance
  • [ ] benches
  • [ ] prompt templates e.g. how X or Y system prompt / template performs on Z or W model. Basically need a list of system prompt and prompt template and inputs and a list of LLM to run test against. Ideally we only run on open source LLM, need high performance maybe, ideally some spare GPUs somewhere

[louis030195]

Few things that could be even better than OpenAI Assistants API but out of scope ATM:

• OpenAI Assistants API assume an assistant always use the same model, but the future will use many different LLMs, for different domain, speed and quality requirements.

[transitive-bullshit]

This project looks awesome 💪

Maybe we can join forces? https://github.com/transitive-bullshit/OpenOpenAI

Prolly the most relevant part of my project is that all of the API routes and input/output types are auto-generated from OpenAI's official OpenAPI spec, so it'll be pretty straightforward to keep them in sync in the future.

[louis030195]

@transitive-bullshit This is cool! Let's chat soon https://cal.com/louis030195/d

[louis030195]

🚢 @transitive-bullshit now using official openai OpenAPI spec