Capture thought signatures from Google models #2305
Conversation
| # thought_signature=_base64_string_to_bytes(item.signature), | ||
| # text=item.content, | ||
| # ) | ||
| # ) |
There was a problem hiding this comment.
NOTE: We don't send ThinkingPart to the providers yet. If you are unsatisfied with this,
Can we enable it here, or is there a broader rule about this in place?
There was a problem hiding this comment.
Yeah we can return the signatures, just not the actual thought text.
| @@ -682,7 +682,7 @@ class ThinkingPart: | |||
| signature: str | None = None | |||
There was a problem hiding this comment.
I suppose the alternative to base64 encoding/decoding would be do signature: str | bytes | None = None, but I'm not sure what kind of ripple effects that would produce.
| @@ -511,7 +525,7 @@ def _process_response_from_parts( | |||
| for part in parts: | |||
| if part.text is not None: | |||
| if part.thought: | |||
| items.append(ThinkingPart(content=part.text)) | |||
| items.append(ThinkingPart(content=part.text, signature=_bytes_to_base64_string(part.thought_signature))) | |||
There was a problem hiding this comment.
Actually, I don't think this is going to work.
The thinking signature in the example (see Appendix 1) comes right after the thinking part.
So, I think support for a signature field would need to be added to ToolCallPart() and TextPart()?
There was a problem hiding this comment.
Hmm, interesting, I had missed that detail in reading your issue originally.
For consistency with Anthropic, and to not have to modify all the other part classes with this Google-specific detail, I would prefer to store the signature on the most recent thinking part in Pydantic AI representation, and then to include it on the (next) text/tool part when sending it back to Google. Can you see if you can make that work?
There was a problem hiding this comment.
Note that we'll also need to handle the signature when streaming, which is implemented here:
Since the signature will come on a text/tool message, we'll want to check if its set there, and if so call handle_thinking_delta(vendor_part_id='thinking', signature=signature) before we call handle_text_delta or handle_tool_call_delta so that the signature will be added to the most recently seen thinking part.
| @@ -511,7 +525,7 @@ def _process_response_from_parts( | |||
| for part in parts: | |||
| if part.text is not None: | |||
| if part.thought: | |||
| items.append(ThinkingPart(content=part.text)) | |||
| items.append(ThinkingPart(content=part.text, signature=_bytes_to_base64_string(part.thought_signature))) | |||
There was a problem hiding this comment.
Note that we'll also need to handle the signature when streaming, which is implemented here:
Since the signature will come on a text/tool message, we'll want to check if its set there, and if so call handle_thinking_delta(vendor_part_id='thinking', signature=signature) before we call handle_text_delta or handle_tool_call_delta so that the signature will be added to the most recently seen thinking part.
Overview
Capture the thought signature for Google Models.
Note: This requires casting the bytes Google provides to a base64-encoded string for compatibility with
ThoughtPart()'ssignaturefield, and then back to bytes when returning the response to Google.Closes #2293