Back to catalog / Agent research

Return traceable citations for document question answering

Provide documents as trusted sources and preserve source locations in answers instead of asking the model to invent citations.

Goal

Answer document questions while linking each checkable claim to an input document.

Prepare a small authorized document set with stable source IDs and recognizable titles/content. [citations]
Follow the citations example using document content blocks with citations.enabled, passing only authorized text in the message. [citations]
Read citation objects and locations, render links to source documents, and verify each span falls within the source. [citations]

Test with a question containing an unsupported trap; citations point only to supplied documents and positions are verifiable.

If no citation is returned, report insufficient evidence; do not fall back to generated page numbers or URLs.

Python 3.11+, the anthropic SDK, a valid API key, and authorized source documents.

Pass only documents the user may read; enforce source access controls before returning results.

This feature depends on supported models and supplied sources; a citation does not prove an inference is correct or authorize disclosure.

These values describe the upstream repository, not popularity or likes for an individual skill or playbook.

No metrics snapshot is available yet.

License: MIT