Citation-Checking Agents
Using search to verify that a claim an LLM made is actually supported by a real, findable source.
How it works
A citation-checking agent runs as a verification pass after a primary answer is generated: it extracts the discrete factual claims from the draft answer, issues a targeted search for each claim, and checks whether the top results actually support (or contradict) the claim rather than just being topically related. This is a narrower, more adversarial search pattern than general RAG — the goal is confirmation or refutation of a specific assertion, not open-ended retrieval.
Example
Draft answer includes the claim 'Company X raised a $50M Series B in 2025.' The checking agent searches "Company X" Series B funding 2025 amount, compares the retrieved figure against the claimed figure, and flags a mismatch if the search results say $40M or a different year rather than silently accepting the original claim.
Pitfalls
- A search returning no results for a claim doesn't prove the claim is false — it may just be under-indexed — so 'unverifiable' needs to be a distinct outcome from 'contradicted,' not collapsed into the same flag.
- Checking a claim against a single search result is fragile if that result itself is wrong or outdated; a robust checker needs at least a couple of independent-looking sources before treating a claim as confirmed.
- Claims that are opinions or predictions rather than verifiable facts get incorrectly flagged as unsupported by a checker that doesn't first classify whether a claim is checkable at all.