FAQ

Q: Why is evidence rejected?
Check GateDecision.reasons (ReasonCode). Common: CTXT_ENTITY_MISMATCH, CTXT_TIME_MISMATCH, CTXT_SOURCE_POLICY_VIOLATION, EVIDENCE_DUPLICATE.

Q: My adapter returns chunks but gating rejects them.
Populate entity_ids, year, provenance.source_type, and (optionally) metadata.doc_type. Use CanonicalFilters.from_state_spec(state) to push constraints into your backend.

Q: How do I control LLM cost/latency?
Use BudgetedProvider (prompt/output caps), RetryingProvider (limited retries), judge budgets (MAX_JUDGE_CHUNKS_PER_CLAIM, MAX_JUDGE_TEXT_LEN).

Q: Can I use my own LLM?
Yes. Implement LLMProviderBase.complete_json and pass to LLMJudge. Decorate with budget/retry wrappers.

Q: How do I see why a verdict happened?
Use TraceBuilder to capture the run; export DOT via TraceGraph.to_dot() and inspect accepted/rejected evidence nodes.

Q: Optional dependencies missing (e.g., qdrant_client/chromadb/boto3)?
Install extras: contextguard[qdrant], contextguard[chroma], contextguard[cloud], contextguard[llm].