aurelio-labs/funkagent

Experimental agent using OpenAI Functions

funkagent's Issues

INFO: I used this as a basis for my Ruby on Rails based agent

See merefield/discourse-chatbot#29 & https://meta.discourse.org/t/discourse-chatbot-supporting-chatgpt/256652/364?u=merefield

Your video and work really got me up to speed quickly on this new Open AI function framework so thank you very much James. 🙏

ISSUE: funkagent should adopt formal function response message format

I've noticed an issue with funkagent, it's not using the formal function response message format described in the documentation

https://platform.openai.com/docs/guides/gpt/function-calling

The format should be

      messages.append(
          {
              "role": "function",
              "name": function_name,
              "content": function_response,
          }
      )  # extend conversation with function response

But instead it's using the assistant role and missing the function name:

funkagent/funkagent/agents.py

Line 82 in a7305d3

res_msg = {'role': 'assistant', 'content': (f"The answer is {result}.")}

This is not critical, it appears, but it's a tidy up and may facilitate dropping the "answer is" text for a tiny token win. And we don't know if this will become more important in the future.

An equivalent change is illustrated below on my Ruby on Rails agent:

merefield/discourse-chatbot@6b89b62

I've PR'd on #9

role function in function response

Bug:
shouldn't https://github.com/aurelio-labs/funkagent/blob/main/funkagent/agents.py#L82 be role 'function' not 'assistant'? This will help the model from not calling the function again with the same query. You expressed this worry on the video and probably this is why it happens.

To add usefulness to functions:
I think you can redact function responses in future questions as only the query/response is really important for context. If the AI needs to know the same thing again it can ask again. (this will limit context window and cut down costs).

It's useful to know if a query is "related" to another query (in the same user context) and if so add previous history otherwise start with 0 history. A semantic lookup on the query/response can be done here at a high threshold. At worsed case you may need to pass the previous response in a secondary check to ensure user is not inferring to the previous response in an ambiguous way.

License?

Hi, thanks for the video and for this cool approach to spinning up an agent! Is the code here open sourced or MIT licensed? I'm considering using it for a project but wasn't sure whether the license is compatible