Generate tool results when using structured result #179
Conversation
| # we have a final result, end the conversation | ||
| result_data = either.data | ||
| # Add all messages to the conversation | ||
| messages.extend(r for r in responses if isinstance(r, _messages.Message)) |
There was a problem hiding this comment.
we extract the Messages from the list 3 times, maybe better to return tuple[_MarkFinalResult[ResultData] | None, list[_messages.Message]]?
Then we don't have to do any extraction aftwards.
|
We should also test if the responses need to be adjacent in the messages, I didn't think they needed to be, but @sudoskys suggested they do. That can be a separate PR. |
|
Looking at those tests now and I like your suggestion about the tuple. Responses do need to be adjacent for OpenAI, not sure if others have the same requirement (message expectations vary). The OpenAI requirement is: a message with one or more tool calls must be followed immediately and only by messages containing responses to every tool call. For example if 2 parallel tool calls are made, the next two messages must be responses corresponding to those two tool call ids. |
|
I think this is GTG. In testing I identified another situation that can result in the same error message, in which the model requests a "regular" tool call and a "final" tool call in parallel in the same payload. Even though this PR adds the |
I think that's what's reported in #161. There's some more discussion needed in that case. |
|
Ok cool, I have a proposal I'll add on that issue. |
|
Thanks so much @jlowin. |
Closes #174