Skip to main content
An answer without evidence is a claim. An answer with citations is a verifiable claim. The Trace shows both — where the evidence came from and how much of the answer is actually grounded in it.

​
Citation view

Every citation in the answer is a superscript number that hovers a preview:
  • Source (document title, memory space name, or “graph fact”).
  • Exact quoted passage.
  • Score (relevance in this retrieval).
  • Click to jump to the source in a side panel.
If your Agent is configured to always cite, you can enforce this — see Grounding.

​
Grounded vs. ungrounded regions

The Trace highlights the answer with two colors:
  • Green — this sentence is backed by a retrieved chunk or graph fact.
  • Gray — this sentence has no direct citation. Could be a bridging sentence, or could be a hallucination.
Fully-ungrounded answers get a warning banner. Not automatically bad — some questions genuinely need reasoning outside the retrieved context — but worth spot-checking.

​
Citation faithfulness

For each cited region, the Trace also shows a faithfulness score — does the cited chunk actually say what the Agent claims it says? Computed by a small verifier model. Scores:
  • 1.0 — perfect match.
  • 0.7-1.0 — supported (paraphrased but consistent).
  • 0.4-0.7 — partial support (parts of the claim are in the source, parts aren’t).
  • < 0.4 — misquote or overreach.
Low faithfulness scores are the strongest signal of a subtle hallucination — the Agent read a real source but drew a stronger conclusion than the source supports.

​
Verified vs. unverified claims

Separate from citations, if the Agent has a causal graph attached, the answer’s factual claims are checked against the graph:
  • Verified — the graph confirms.
  • Contradicted — the graph disagrees.
  • Unchecked — no matching variable or edge in the graph.
Shown next to each factual claim. See Verify agent answers.

​
Missing citations

If the Agent stated a fact but didn’t cite it — and retrieval had a relevant chunk it could have cited — the Trace flags it: “Missing citation: ‘Refund policy is 30 days’ — matching chunk: policy-2026.pdf#refunds.” Turn into a Signal to improve future prompts.

​
Citation formats

The default format is superscript numbers. Configurable per Agent to:
  • Inline URLs.
  • Footnote-style at end of answer.
  • Structured (JSON) — for programmatic consumers.
Pick a format that matches how your users consume the answer.

​
Debugging “why isn’t this cited”

If you know a source has the answer but the Agent didn’t cite it, work back through the trace:
  1. Did retrieval return this source? See Diagnostics.
  2. Did it rank high enough to be in the top-K?
  3. Did the Agent decide it was relevant enough to cite?
The Trace shows each stage so you can locate the specific point where the source dropped out.