mistral: read tool calls from AIMessage #20554
Conversation
dosubot
bot
added
the
size:M
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎ 1 Ignored Deployment
|
dosubot
bot
added
the
🤖:improvement
| @@ -216,8 +247,15 @@ def _convert_message_to_mistral_chat_message( | |||
| if _id := tc.get("id"): | |||
| chunk["id"] = _id | |||
| tool_calls.append(chunk) | |||
| elif message.tool_calls or message.invalid_tool_calls: | |||
There was a problem hiding this comment.
Should this be prioritized over additional_kwargs? The latter should be the source of truth if swapping providers etc
There was a problem hiding this comment.
I'm open to either. I actually implemented it that way first, but it introduces small issues with casting to/from Mistral dicts, since the additional_kwargs contain (1) information on order of tool calls (we lose that with the separation into valid and invalid tool call lists), and (2) JSON formatting, since we parse and then dumps the args. Our unit tests maintain that we can go back/forth so I wasn't sure if we wanted to lose that. These are small issues; if we think there's a compensating benefit to prioritizing .tool_calls then happy to change it.
There was a problem hiding this comment.
ah yeah if we have a provider that (1) formats tool_calls into additional_kwargs, and (2) does not use OpenAI-style formatting then this will break. updated this. see changes to unit tests.
There was a problem hiding this comment.
@efriis see above - is it OK to lose the behavior maintained in unit tests, in which applying _convert_message_to_mistral_chat_message and _convert_mistral_chat_message_to_message returns the original message identically?
dosubot
bot
added
the
lgtm
ccurme
had a problem deploying
to
Scheduled testing
GitHub Actions
Failure
ccurme
had a problem deploying
to
Scheduled testing
GitHub Actions
Failure
Co-authored-by: Eugene Yurtsev <[email protected]>
No description provided.