Answer Verification

Answer Verification is enabled by default for every crate. After an answer is generated, ContextCrate sends the answer and the retrieved source chunks to the configured answer model. The model must reply yes only when every factual statement is supported by those chunks.

Verification buffers the generated answer until the check finishes, so answers no longer arrive token-by-token. Retrieved sources are still sent before the final answer. Configure the feature in the crate's RAG settings and choose one action when unsupported claims are found:

• Revise once (default) asks the model to rewrite the answer using only supported source facts.
• Block answer returns the standard knowledge-base no-answer response.
• Return with warning returns the original answer with an unsupported-claims warning.

If verification fails or produces any response other than yes or no, ContextCrate returns the original answer with a not-verified warning. The answer stream emits a verification event with verified, revised, blocked, unsupported, or unavailable status before answer text.

Privacy and limitations

Verification uses the configured OpenAI-compatible answer endpoint, model, credentials, headers, and timeout. It sends retrieved chunk content and generated answer text to that provider, adds at least one LLM request and latency, and is not a guarantee of factual correctness. Prompts, chunks, answers, and verification results are not written to audit logs. Use TLS and treat the provider as a data processor.