As 0.10.11 was released yesterday, is it

any 0.10.x version should be compatible with another 0.10.x version -- breaking changes happen for example 0.10.x -> 0.11.x

so safe to update!

Hm, ok trying to proove that, currently strange things happen.

It claims using tools by "Action: ... but does not seem to

Reading documentation...

My impression is, that for example tool description are somehow part of a prompt? Is that correct? And should help the agent to select that tool?

The thing is, the documentation seems to be quite good and also up2date but this concepts are not explained. Or I did not find it yet.

Depending on the LLM, it may be hallucinating the tool response :PSadge:

Yes I guess, you mentioned it already. Thanks for that. But actually that "missing prompting documentation" should somehow be addresses. Basically three issues:

• Is the tool description used as prompt and best practices what to put there
• I try to icecream agent.get_prompts, this is always empty? Is this correct?
• Sometimes I get "Could not parse output: Thought: ..... " what might be the cause of this?

Ok, thanks for 1. and 2.

1. might be some internal issue? How can this be debugged?

I mean, its mostly an issue with the LLM outputting an unexpected format, that can't easily be parsed

Try adding this to the top of your code, it will print every llm input and output

Thanks

Oh wow, lot's of stuff, cool 😃

According to what I see, that "Could not parse output..." is likely due to broken json.

I have one more question and would be very happy if you could answer it:
So, my understanding is like this: Let's say i have a simple document store with one vectorized document which are in fact embeddings. Now my query is tokenized and this is converted into embeddings. Now there is (my understanding) a "cosine similarity" or something similar search for some "hit" in the vector store.
Now the important question: How is that "hit", means the "matching" embedddings from the vector store fed into the llm?

The top-k matching results (calculated with cosine similarity or similar) are retrieved, and given to a response synthesizer. That response synthesizer puts the retrieved nodes + user query into a prompt template, and asks the LLM

Thanks, I think I understand it now (while looking at the json stuff in my vectorstore)