search

brainlid / langchain

Elixir implementation of an AI focused LangChain style framework.
https://hexdocs.pm/langchain/
Other
571 stars 68 forks source link

Handle data extraction where tool_calls response is a list. #127

Closed cpursley closed 3 months ago

cpursley commented 4 months ago

A high percentage of the responses for the data extraction chain have been returning a list of tool_calls instead of a single result in the "info". This PR maps "info" when there are multiple tool calls. Not sure how to set up a test case for that but open to ideas.

brainlid commented 4 months ago

@cpursley yes, the data extraction chain needs to be fixed up. I'm working through a big change right now that will make the data extraction process easier particularly for models that don't support tool calls (like Bumblebee self-hosted models).

Because every model is so unique in terms of what it can do and the prompts needed for a successful extraction, I'm less and less convinced that a standard one-size-fits-all chain can work well. It might be more of, "Here's a guide and a template for building your own data extraction process".

I'm hopefully finishing up the code changes now and will write more about this.

cpursley commented 4 months ago

This also might be that I'm using Llama but via an OpenAI compatible API fwiw.

Sounds good, let me know how I can help.

brainlid commented 3 months ago

This may still be an issue with the new version. I'm not sure. You'll have to give it a try and let me know what's happening.

cpursley commented 3 months ago

This seems to have resolved itself with the new version. Closing.