Citation Quality
Your AI shows exactly which document each statement came from — and says so honestly when a statement has no source — so people can check its work instead of just trusting it.
Think of a librarian double-checking a student's essay. Every claim gets a note: this fact came from that page of that book. And when a claim has no book behind it, the librarian flags it rather than letting it slide. We make the AI's answers pass that librarian.
The problem, in plain words
Your AI writes a confident, fluent paragraph. It sounds right. Is it? Maybe there's a source link at the bottom — but does that source actually say what the paragraph says? An answer that cites the wrong document is worse than one that admits it doesn't know, because it looks trustworthy while being wrong. This is the famous hallucination problem: smooth text with nothing behind it. And when there is no trail from each claim back to a source, nobody can check a single sentence without redoing the entire research themselves.
What we set up
Three layers. First, retrieval hands the model passages that carry stable IDs — every chunk of text has a permanent label, like a page number that never changes, so it can always be pointed back to. Second, the model is instructed to attach every factual claim in its answer to one of those passages. Third, a checker — sometimes simple rules, sometimes another model acting as verifier — reads behind the AI: does the cited passage actually support the claim? Claims with no support get flagged instead of shipped. And every failure the checker catches feeds back into improving the prompt or the retrieval, so the same mistake gets rarer.
How it works, step by step
- Sources come with permanent labels
Every retrieved passage has a stable ID, so a citation always points somewhere real and findable.
- Every claim points at a passage
The model must attach each factual statement to its source — showing its work, sentence by sentence.
- A checker reads behind the AI
Claim by claim, it verifies that the cited passage really says what the answer says it says.
- No source? The claim is flagged
An unsupported statement is marked as unsupported — not dressed up as solid fact.
- Failures make the system better
What the checker catches feeds back into prompt and retrieval fixes, so the unsupported-claim rate trends down.
What changes for you
Before: trust was all-or-nothing — a reader either believed the answer or redid the research from scratch. After: every claim is checkable in one click, and the rate of unsupported claims becomes a number your team tracks and pushes down. Hallucinations stop being invisible; they show up as flags you can count. What it won't do: it cannot make a wrong source right. If your documents contain an error, the AI will cite that error faithfully — citations prove where an answer came from, not that the origin itself is correct.