Step 4 is Missing: How to feed tool results back into chat history
#28
by
Cyleux
- opened
Please
From the perspective of the model right now, after we do the second generation after the tool call to interpret the tool response, if we feed just that back into the chat history, it appears to the model that it hallucinated / generated it rather than doing a tool call.
Then when this happens a couple times the tool calling layer learns to stop calling tools and just calls respond every time and the model at that point (after a chat history where it appears to directly respond) is happy to hallucinate whatever.
In my mind, there is missing format template function for grounded tool use.
Thank you
Cyleux
changed discussion status to
closed
I have the same question, do you have any solutions?