A prompt grounding verifier answers the question that matters most in retrieval-augmented and document-Q&A systems: did the model actually stick to the context I gave it, or did it make something up? Hallucinations are rarely obvious — they are fluent, confident sentences that the source never supported. This tool compares a model’s answer against the context you provided and flags sentences whose key terms cannot be traced back, so you know exactly which claims to verify.
How it works
You paste two things: the context (the source text or retrieved documents the model was given) and the answer the model produced. The verifier splits the answer into sentences, extracts the meaningful key terms from each (dropping common stop-words), and measures what fraction of those terms appear in the context. Sentences with strong overlap are marked grounded; those with weak overlap are flagged as potentially unsupported. It reports an overall grounding score and highlights the suspect sentences. All of this runs locally in your browser — nothing is uploaded — so it is safe for proprietary material.
When grounding matters most
Grounding is not equally important in every LLM application. It is most critical in systems where the model is expected to report from provided evidence rather than reason from general knowledge:
- Retrieval-augmented generation (RAG): The whole premise is that answers come from retrieved documents, not model training. Ungrounded sentences are by definition hallucinated.
- Document Q&A: A user asks about a specific contract, report, or policy. Answers that blend the document with general LLM knowledge can be wrong in ways that are legally or operationally significant.
- Summarization with citations: Any summary that claims to represent a source must be verifiable against it.
Grounding matters less in creative tasks, brainstorming, or general-knowledge questions where no specific source was provided.
Understanding the grounding score
The verifier reports a per-sentence overlap score and an overall grounding percentage. How to read these:
- High overlap (strong grounding): The sentence’s key terms appear directly in the context. This is a good signal but not proof — a model can copy terms while distorting meaning.
- Low overlap (flagged): The key terms do not trace back to the context. This is the primary hallucination signal. Verify the claim manually.
- Overall score: The fraction of answer sentences that are well-grounded. A RAG system targeting high faithfulness should aim for a consistently high score across many answer/context pairs, not just one.
The score is heuristic, not semantic. Two common false signals:
- Over-flagging: A correct paraphrase uses different vocabulary from the source, so it scores as low-overlap even though it is accurate.
- Under-flagging: A claim copies the context’s vocabulary but changes a number, date, or named entity — the verifier misses this.
Both are reasons to treat flagged sentences as “needs human review,” not “definitely wrong.”
Improving faithfulness upstream
The verifier tells you where grounding failed. Fixing it upstream — in the prompt — is more durable than reviewing outputs case by case:
- Instruct explicitly: “Answer only using information from the provided context. If the answer is not in the context, say ‘The context does not contain this information.’”
- Ask for source spans: “For each claim, quote the supporting passage from the context.” Run the verifier on the quoted version — it will be trivially grounded.
- Penalize speculation: “Do not infer, extrapolate, or draw on general knowledge. Only report what the context explicitly states.”
Use the verifier after each prompt change to confirm that faithfulness improved, not just that the answer sounds better.
Tips and examples
This is a triage tool: its job is to point you at the sentences worth reading carefully, not to deliver a final verdict. Word overlap sometimes over-flags a correct paraphrase and can miss a claim that borrows the context’s vocabulary while distorting its meaning, so always confirm flagged sentences by hand. For audit trails, ask the model to quote the supporting span for each claim, then run the verifier on the result. Use it after every change to a RAG prompt to confirm faithfulness did not regress.