Anthropic Citations (Messages API)

Type: official documentation Tier: 1 (Official Docs) Author(s): Anthropic Date: accessed 2026-06-01 URL: https://docs.anthropic.com/en/docs/build-with-claude/citations Accessed: 2026-06-01

Why This Source Matters

RAG's promise is not just "answer from documents" but "answer from documents and show me where." Project 04 builds a citation system that traces each claim back to the source passage. This doc is the primary source for how a production API does exactly that — mapping spans of the answer to exact character/page/block locations in the source documents — and is the model for the (provider-agnostic) citation contract the learner builds. It grounds the lesson's citations section and the "answers must be verifiable" stake.

Key Claims

What citations do

• Claude "is capable of providing detailed citations when answering questions about documents, helping you track and verify information sources in responses."
• Each answer text block "can contain a claim that Claude is making and a list of citations that support the claim." Citations "reference specific locations in source documents."

How it works (three steps)

1. Provide documents, enable citations. Attach documents (plain text, PDF, or custom content) with citations.enabled = true.
2. Documents get chunked. "Document contents are 'chunked' in order to define the minimum granularity of possible citations." Plain text and PDFs are "chunked into sentences"; custom content blocks "are used as-is and no further chunking is done." (Note: the citation granularity is sentence-level chunking — distinct from, but conceptually the same operation as, the retrieval chunking in sources/articles/chunking-strategies.md.)
3. Claude returns a cited response. Answer is split into multiple text blocks; cited blocks carry a citations list pointing into the source.

Citation format — answer span → source location

• Plain text → char_location with start_char_index / end_char_index (0-indexed, exclusive end).
• PDF → page_location with start_page_number / end_page_number (1-indexed, exclusive end).
• Custom content → content_block_location with start_block_index / end_block_index (0-indexed).
• Each citation carries cited_text ("The exact text being cited") and a document_index (0-indexed across all documents in the request).
• cited_text "is provided for convenience and does not count towards output tokens" — and "is guaranteed to contain valid pointers to the provided documents."

Citable vs. non-citable content

• Only text in a document's source "can be cited from." Optional title and context fields "will be passed to the model but not used towards cited content" — context is a place for document metadata.

Why the feature beats prompt-only "please quote the source"

• "Better citation reliability: ... citations are guaranteed to contain valid pointers to the provided documents." (Prompt-asked quotes can be subtly wrong or fabricated.)
• "Improved citation quality: ... significantly more likely to cite the most relevant quotes ... as compared to purely prompt-based approaches."

Relevant To

• concepts: [citations, grounding, provenance, attribution, hallucination-detection, document-qa, verifiability]
• projects: [04-pdf-research-assistant, 07-ai-evaluation-framework]

Notes

• Two kinds of "chunking" live in a RAG system and learners conflate them. (1) Retrieval chunking — how you split a PDF into vectors for the vector DB (sources/articles/chunking-strategies.md). (2) Citation chunking — the sentence-level granularity at which an answer span maps back to source text (this doc). Project 04 uses the same chunk ids for both: retrieve a chunk, then cite the chunk it came from.
• The portable lesson (provider-agnostic, since the lab defaults to Groq/Ollama via LiteLLM): a citation is a pointer from a span of the answer to a (document, chunk/char/page range, exact quoted text). The learner builds this by keeping each retrieved chunk's id/source and having the generator attach the supporting chunk id to each claim.
• Caveat learners hit: citations are incompatible with strict Structured Outputs in this API, because "citations require interleaving citation blocks with text output." If you force a rigid JSON schema you lose inline citations — a real design tension.
• Pairs with sources/papers/ragas.md: citations make faithfulness checkable — if a claim cites chunk 7, you can verify the claim against chunk 7's text.