What is Citation?
Updated on
Citation is a verifiable mapping from an AI-generated claim to a specific source span, record, or tool result intended to support that claim, together with enough provenance to retrieve and inspect the evidence.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Represent provenance, not just a URL
A citation record should carry a claim or sentence ID, stable source ID, source version, page or section, character or token offsets when available, a display label, and the retrieval authorization context. Store the quoted span or a tamper-evident reference when sources can change. The renderer must preserve the mapping after answer streaming, sentence edits, deduplication, and document reindexing.
Measure support and coverage separately
Citation support or correctness evaluates whether each cited source actually entails its claim. Citation coverage or completeness evaluates whether claims that require external evidence have adequate supporting citations. A response can score high on one and low on the other: every attached citation may be correct while half the factual claims remain uncited, or every claim may have a link while several links fail to support it.
Design for verification and access control
Prefer claim-level or sentence-level anchors when users must verify decisions, but retain surrounding context so a quote is not misleading. Reject stale offsets, missing versions, inaccessible sources, and citations introduced only after generation without entailment checks. Never reveal the existence, title, snippet, or URL of a source the current user cannot access; citation authorization must match retrieval authorization.
Key Characteristics
- Maps a specific answer claim to inspectable evidence
- Preserves source identity, version, location, and display metadata
- Separates citation support correctness from claim coverage
- Requires stable anchors across streaming, editing, and reindexing
- Does not by itself prove source authority, freshness, or factual truth
- Must enforce the same access policy as retrieval and answer generation
Common Use Cases
- Showing evidence spans for enterprise policy answers
- Tracing agent conclusions back to database rows or tool outputs
- Auditing medical, legal, financial, and compliance claims
- Measuring citation support and coverage in RAG regression tests
- Debugging stale, broken, overly broad, or unauthorized source links
Example
Loading code...Frequently Asked Questions
What makes a citation correct in a RAG answer?
The cited span must support the exact claim under the stated qualifiers, numbers, and time scope. Topical similarity or sharing keywords is insufficient. Reviewers should be able to open the source version and verify the entailment without searching an entire document.
What is the difference between citation support and citation coverage?
Support measures whether citations that are present actually justify their claims. Coverage measures whether all claims that require external evidence have adequate citations. Report both because a system can attach a few perfect citations while leaving many factual claims unsupported.
Is a document-level citation sufficient?
It may be acceptable for short, stable documents, but passage, section, page, or record-level anchors reduce verification cost and ambiguity. High-stakes workflows should retain the exact source version and enough surrounding context to detect misleading quotations.
How should citations work when documents change?
Use immutable source versions or snapshots and store stable IDs plus offsets or structural anchors. When reindexing invalidates a mapping, mark the citation stale and regenerate or suppress it rather than silently pointing to changed content.
Can a citation expose restricted information?
Yes. Titles, snippets, URLs, and even the existence of a document can be sensitive. Apply authorization before retrieval and again before rendering citations, and avoid persisting protected evidence in logs available to broader audiences.