Search

Answer a question from memory, with citations

Corrective RAG over the same retrieval stack as /v1/search: generate an answer, then verify each claim against the facts it cites. The pipeline ABSTAINS rather than guess — answer is null and reason names the gate that fired (no results, thin coverage, ungrounded support, a failed verifier…). synthesisGuardrails picks the policy: strict closes to null on a partial verdict, lenient returns the answer with the verdict attached, off skips the verifier. Cited facts are always the caller’s own visible memory. Fans out to the shared embedding / LLM budget, so it sits in the per-credential expensive throttle bucket — 10 requests/min, not the 120/min default. Source: src/synthesize/synthesize.controller.ts.

Required scope: brain:read.

post/v1/synthesize

Request

  • Base URL: https://brain.inite.ai
  • URL: https://brain.inite.ai/v1/synthesize
  • Auth: HTTP bearer

Request body

answerLangstring
asOfstring
confidenceFloornumber
disableLangFilterboolean
entityIdsstring[]
entityTypesstring[]
explainboolean
includeContestedboolean
includeRetractedboolean
includeStaleboolean
limitnumber
minConfidencenumber
outputShape'full' | 'compact' | 'ids'
predicatesstring[]
querystring required
queryLangstring
requireProvenanceboolean
searchMode'vector' | 'lexical' | 'hybrid'
synthesisGuardrails'strict' | 'lenient' | 'off' | 'answer'
synthesisModelstring
tokenBudgetnumber
userIdstring

Response

The answer (or a reasoned abstention).

answerstring nullable required
cachedboolean
decisionLogDecisionLogEntry[]
reason'no_results' | 'no_grounded_evidence' | 'low_coverage' | 'evidence_capability_unmet' | 'ungrounded_evidence' | 'verifier_failed' | 'verifier_partial' | 'generator_error' | 'verifier_error'

Changes

No recorded changes to this operation across all 1 revision of this API.