[Tracking][WebLLM] Function calling (beta) and Embeddings

This issue tracks various action items we would like to complete with regard to the features function calling and embeddings.

Function calling (beta)

We are calling it beta because multiple iterations may be needed for function calling. It may be hard to conform different open-source models' function calling formats to OpenAI API. We will try to make each iteration non-breaking.

[x] F1 Enable manual function calling (completed via https://github.com/mlc-ai/web-llm/pull/527, supported in npm 0.2.53)
- That is, function calling with only system, user, assistant, tool messages, without using the tools and tool_calls fields of OpenAI API
- Reach parity with the examples provided in MLC-LLM: https://github.com/mlc-ai/mlc-llm/pull/2744
- This requires various runtime changes as initiated in https://github.com/mlc-ai/web-llm/pull/467
- Add examples for Llama3.1 and Hermes2 after supported
[ ] F2 Support function calling following OpenAI API with tools and tool_calls
- Essentially supporting the example in https://platform.openai.com/docs/guides/function-calling.
- There are new fields in the official OpenAI API, which we should support as well if possible.
- This may limit the flexibility for the user. For instance, while Llama3.1 offers roughly 3 formats for function calling, using tools will force us to use only one of them
- The previous PR only offers minimal one-round function calling support https://github.com/mlc-ai/web-llm/pull/451
- We want to allow the model to make tool calls or respond in natural language at its own discretion
[ ] F3 Use BNFGrammar to guarantee tool call generation correctness
- This requires the model to use a special token to signify the beginning of a function call, <tool_call> in the case of Hermes2. Upon such a token being generated, we instantiate a BNFGrammar instance. When ended, force it to generate </tool_call>. Before and after this tool call, the model can either generate natural language or other tool calls.
- The previous PR forces the response to be a tool call, which limits the flexibility a lot https://github.com/mlc-ai/web-llm/pull/451
Related issues
- https://github.com/mlc-ai/web-llm/issues/462
- https://github.com/mlc-ai/web-llm/issues/297

Embedding, Multi-model Engine, Concurrency

[x] E1 Loading multiple models within an engine (completed via https://github.com/mlc-ai/web-llm/pull/542, supported in npm 0.2.59)
- For applications like RAG, two models are needed to complete this, one embedding model and one LLM. We'd like to hold all models in a single MLCEngine instead of instantiating multiple engines. This makes MLCEngine behave like an endpoint, and offers the possibility for intra-engine optimizations in the future.
- Each model can process requests concurrently if needed: https://github.com/mlc-ai/web-llm/pull/546 (published in npm 0.2.60)
[x] E2 Fix concurrent request issues (completed via https://github.com/mlc-ai/web-llm/pull/549, supported in npm 0.2.61)
- With a single model, we encounter correctness issues with multiple concurrent requests as broughtup in https://github.com/mlc-ai/web-llm/issues/522
- After E1, we need to further pay close attention to potential concurrency issues.
[x] E3 Implement engine.embeddings.create() (completed via https://github.com/mlc-ai/web-llm/pull/538, supported in npm 0.2.58)
[x] E4 Add an example for RAG (completed via https://github.com/mlc-ai/web-llm/pull/550)
Related issues

mlc-ai / web-llm

[Tracking][WebLLM] Function calling (beta) and Embeddings #526

Function calling (beta)

Embedding, Multi-model Engine, Concurrency