Handbook

Task `answer_from_evidence` v1

Answer a question using only the structured evidence object; return answer, calibrated confidence, citations to evidence ids, and boolean / list flags when context is insufficient.

Updated

Task answer_from_evidence v1

Summary

Answer a question using only the structured evidence object; return answer, calibrated confidence, citations to evidence ids, and boolean / list flags when context is insufficient.

Inputs

Field Type Required Notes
question string yes User question (non-empty after strip for production paths).
evidence object yes Evidence pack (may be {} when unknown). Keys are evidence ids.
temperature number no Default 0.05.
extra_body object no Optional extra envelope fields for transport.
canonical_user_json boolean no Emit canonical JSON when true.

Output

Top-level JSON (see contract.json output_schema):

Field Type Notes
answer string Final answer text.
confidence number In [0.0, 1.0] after coercion.
citations array Structured citations with evidence references.
insufficient_evidence boolean True when answer should not be relied on.
missing_evidence array Ids or hints for missing material.

Examples

{
  "question": "What policy applies?",
  "evidence": {
    "e1": { "text": "MFA is required for all staff." }
  }
}

Failure modes

  • SchemaFailure — evidence not an object, or output JSON missing required keys.
  • ParseFailure — assistant returned non-JSON or wrong shape.
  • GatewayFailure — upstream error envelope.

Verification

Task-internal JSON verification and rag.citations when enabled in verification_policy (see contract.json).

Changelog

  • v1 — Initial evidence-backed answer task.