feat(mistralai): surface citation metadata from chat responses

[JulienRabault]

Fixes #36427

Changes

• Extract Mistral reference content chunks into AIMessage.response_metadata["citations"].
• Preserve non-streaming AIMessage.content as a string when Mistral returns typed text / reference chunks.
• Extract reference chunks from streaming delta.content into chunk response_metadata["citations"].
• Keep streaming content unchanged; no content_blocks conversion or citation offset calculation is added.

Why

Mistral citation responses can return content as typed chunks, including reference chunks with reference_ids.

Before this change, ChatMistralAI kept only message content and dropped the citation metadata. RAG users could not map Mistral answers back to source documents through LangChain.

References:

• https://docs.mistral.ai/capabilities/citations/
• https://docs.mistral.ai/resources/cookbooks/mistral-rag-mistral-reference-rag

Review notes

• The raw Mistral reference chunks are stored in response_metadata["citations"].
• This PR does not convert citations into v1 content_blocks annotations.
• This PR does not compute citation offsets.

AI assistance disclosure: AI tooling helped prepare this change; the submitted diff was reviewed and tested before submission.

[ccurme (ccurme)]

The ideal format for citations is:

• Mistral reference blocks in .content (untyped)
• add translation to and from the standard langchain TextContentBlock in langchain_mistralai/_compat.py. This will translate the reference blocks into a text content block with an Annotation.

If it's possible to do this in a non-breaking way (e.g., if users need to opt into receiving references), this is preferred. It might be OK to make this change in a 1.2 release.

Otherwise, we can add to additional_kwargs (which is a more typical place for provider message data than response_metadata). The issue with this is we lose the sequencing of the reference in the text blocks.

[JulienRabault]

Thanks, that makes sense.

I held off on converting the Mistral reference blocks into annotations because I don't see text offsets in the provider response, and I don't want to invent positions.

I'll update the PR to preserve the native reference blocks in .content and add the _compat.py translation to standard TextContentBlock / Annotation.

If keeping those blocks in .content is too breaking for this release, I can move the provider-native data to additional_kwargs instead. Do you want me to take that fallback path here?

[JulienRabault]

any news ?