ChatMistralAI: Surface Citation Metadata from Chat Responses

What it does

ChatMistralAI now surfaces citation metadata from chat responses. When Mistral returns inline citations referencing source chunks used to ground an answer, LangChain exposes them as structured metadata on the AIMessage via additional_kwargs["citations"].

Why it matters

In RAG pipelines, citations bridge the LLM output and the documents that informed it. Without them you get answers but no accountability. With them you can trace answers to source chunks, highlight passages in a UI, filter low-confidence retrievals, and build trust in production.

Example

$ Ask a RAG-style question and extract citations from the response.

=== Answer ===
The 2024 partnership agreement established co-development terms for the joint venture,
including a 50/50 revenue split and a three-year exclusivity clause.

=== Citations (2 found) ===
  [1] {"document": "partnership_2024_v1.pdf", "start_index": 1247, "end_index": 1380}
  [2] {"document": "partnership_2024_v1.pdf", "start_index": 2105, "end_index": 2230}

Common flags

ChatMistralAI.invoke()

Returns citations in additional_kwargs

ChatMistralAI.bind_tools()

Bind tools/schema to force structured outputs

AIMessage.additional_kwargs

Access raw provider metadata on any AIMessage

with_structured_output()

Get type-safe structured responses with a Pydantic/Zod schema