community: Add "headers" parameter support to OpenAPI tools

[unkindypie]

• Description: Allow OpenAPI agent to use request headers.
  It adds the following:
  • request tool now supports passing of the headers parameter. Headers are optional not to introduce breaking changes. Request headers are merged with headers passed to RequestWrapper class overwriting duplicate names.
  • OpenAPI planner prompts were updated to support new parameter.
  • reduce_openapi_spec now has new optional parameter "remove_optional", which allows to control whether optional parameters will be added to reduced API specification. This is required for the API I am integrating langchain with and often needed in real world scenarios.
  • new functionality for requests tool covered with tests
  • docstrings updated

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Jan 3, 2025 3:41pm

[unkindypie]

Example repo which uses new features, although client code won't differ from the current version except for new parameter in the reduce_openapi_spec function https://github.com/unkindypie/langchain-playground/blob/master/openapi_agent.py

[ccurme]

Thanks for this. I'm wary of the current implementation because all LLMs currently using the requests tools will need to begin populating (empty) headers in their tool calls.

Is it desirable for the LLM to generate headers? Curious about the context. If they can be programmatically generated outside the LLM, we have a guide on passing those to tools here.

[ccurme]

Does defaulting to False change behavior for existing users?

Currently we are effectively removing optional by checking parameter.get("required"), right?

[unkindypie]

No, False will behave as it was working before my changes and True will keep the optional params which is really useful in my case

Currently we are effectively removing optional by checking parameter.get("required"), right?

Right, we just remove anything which is not required at the moment

[unkindypie]

Thanks for this. I'm wary of the current implementation because all LLMs currently using the requests tools will need to begin populating (empty) headers in their tool calls.

Yes, but in theory nothing should change as the description of the tool is updated. I've tried on my end and it worked on GPT 4.0

Is it desirable for the LLM to generate headers? Curious about the context. If they can be programmatically generated outside the LLM, we have a guide on passing those to tools here.

It's desirable in my case because headers must be passed only in a certain case based on description of a header which an LLM should determine, here is an example from our Swagger, but thank you for sharing.

After my changes to the package works quite cool, LLM reads from what endpoint it needs to get ID based on a description from an optional header, adds it to the path and makes an additional request

[unkindypie]

@ccurme Also regarding filling out empty headers problem, for this purpose I've made it default to {}, so there must be no changes required for the existing code and also LLMs can just omit and don't pass it, and the code will still work