Answer Verification
Flag the answer that has no source.
Answers built from the knowledge base are checked against the document they rest on. When a claim is unsupported, the answer is either flagged or withheld.
- Verdicts
- 3
- Judges
- 2
- Default
- off
- Confidence
- 0.0 - 1.0
The decision
With no source, the model is not let loose.
The grounding step asks two questions and takes one of three paths. The third is the one that matters: when a question needs a source and none was found, the chain closes rather than opens.
The user's question
Does it need a source?
A question the knowledge base is not needed for; the model runs normally.
Was a source found?
The answer may rest only on the retrieved chunks, and is verified against those same chunks.
The model is not let loose on this question. No answer beats a quiet invention.
Verdicts
Three outcomes.
Each verdict returns with a one-sentence reason and a confidence between 0 and 1.
Every claim in the answer is supported by the context.
An unsupported claim is present; the answer returns but is flagged.
The claim contradicts the context; the answer is withheld.
Judge
Two ways to evaluate.
With no model judge available the system falls back to the lexical path; verification degrades rather than disappearing.
llm
Model judge
Gemini or a local model reads the context and the answer, returning a verdict, a reason and a confidence.
lexical
Lexical fallback
With no judge available, Jaccard overlap is measured against a threshold; no model call needed.
Why is it off by default?
The model judge is a second LLM call and roughly doubles unit cost. So the decision is yours: it is enabled per policy, on the flows where correctness matters more than cost.
See it with your own data.
In a technical session we run your prompts and your policy through it live.