Citation & Attribution
The agent explicitly cites the sources — documents, tool outputs, retrieved passages, URLs — that support its claims. Output includes references linking assertions to their evidence. This builds trust, enables verification, and reduces the impact of hallucination by making the evidence chain visible.
Structure
The agent generates a response with inline citations referencing specific source documents. An optional verification step checks that cited claims are actually supported by the referenced sources.
How It Works
- Retrieve — gather relevant source documents (via RAG, search, or tool calls)
- Generate with citations — agent produces response with references to specific sources
- Format references — citations are formatted inline (footnotes, brackets, links)
- Verify grounding — optionally check that each cited claim is supported by the referenced source
- Flag ungrounded claims — remove or mark any assertion not supported by sources
Citation formats:
- Inline footnotes —
"The policy was updated in March [1]."with source list - Source metadata — each response chunk carries its source document ID
- Confidence indicators — flag claims with vs. without source support
- Direct quotes — include verbatim excerpts from sources
Key Characteristics
- Verifiable — users can check the source behind any claim
- Trust building — cited responses are more trustworthy than uncited ones
- Hallucination visible — unsupported claims are easier to spot when citations are expected
- Source quality matters — citing bad sources doesn't help
- Overhead — citation generation and verification add complexity and tokens
When to Use
- Accuracy matters and users need to verify claims (research, legal, medical, finance)
- You're building RAG systems where responses should be traceable to source documents
- Trust and transparency are product requirements
- Regulatory or compliance contexts require evidence for assertions
- Users have asked "where did you get that?" more than once