What is Citation?
Citation is the attribution of an AI-generated claim to a specific source document, passage, URL, record, or tool output that supports it.
How It Works
Citation is the user-visible part of grounding. In RAG systems, citations help readers verify claims, inspect source context, and detect unsupported answers. Strong citations point to the specific passage or record that supports a statement, not merely to a broad document. They also preserve source metadata such as URL, title, timestamp, page, section, and access scope. Citation quality depends on retrieval quality, chunk boundaries, source offsets, and whether the generation step actually uses the cited evidence.
Key Characteristics
- Attributes generated claims to concrete sources or evidence records
- Improves auditability and user trust when citations are specific and correct
- Requires source metadata, chunk offsets, and stable document identifiers
- Can be misleading if the cited source does not actually support the claim
- Works with documents, web pages, database rows, tool outputs, and uploaded files
Common Use Cases
- Showing source passages under a RAG answer
- Linking policy answers to exact handbook sections
- Tracing agent decisions to tool outputs and retrieved records
- Helping users verify medical, legal, financial, or compliance claims
- Debugging unsupported answers by inspecting cited evidence
Example
Loading code...Frequently Asked Questions
What makes a good citation in RAG?
A good citation points to a specific source span that directly supports the claim, with enough metadata for a user or auditor to verify it.
Can citations be wrong?
Yes. A system can attach a citation to a source that is related but does not support the exact claim, so citation faithfulness must be evaluated.
Is a document-level citation enough?
Sometimes, but passage-level or section-level citations are usually better because they reduce the effort needed to verify the answer.
How do citations relate to chunking?
Chunking determines source spans and offsets. Poor chunking can make citations too broad, incomplete, or hard to audit.