Interface RecallResult

Type declaration

• candidatesScanned: number

• vectorSearchTimeMs: number

• scoringTimeMs: number

• totalTimeMs: number

• Optional policyProfile?: MemoryRetrievalProfile

• Optional suppressed?: "weak_hits"

• Optional confidence?: RetrievalConfidenceSummary

• Optional escalations?: string[]

• Optional hyde?: {
      hypothesis: string;
  }

  Step-4: when a HybridRetriever runs with a hydeRetriever attached, the first ~120 chars of the generated hypothesis are surfaced here for post-hoc analysis of which queries benefited from expansion.

  • hypothesis: string

• Optional factSupersession?: {
      droppedIds: string[];
      parseOk: boolean;
      llmLatencyMs: number;
      notes?: string[];
  }

  Step-5: post-retrieve FactSupersession pass diagnostics. Populated only when a bench adapter or downstream consumer ran FactSupersession.resolve() over the retrieved traces.

  • droppedIds: string[]

  • parseOk: boolean

  • llmLatencyMs: number

  • Optional notes?: string[]

• Optional splitOnAmbiguous?: {
      threshold: number;
      candidateCount: number;
      replacedIds: string[];
  }

  Step-6: when HybridRetriever runs with splitAmbiguousThreshold set, the bottom fraction of traces by first-pass rerank score are split at sentence boundaries and rescored. Replacements are recorded here for post-hoc analysis.

  • threshold: number

  • candidateCount: number

  • replacedIds: string[]

• Optional stageIds?: {
      dense: string[];
      sparse: string[];
      merged: string[];
      reranked: string[];
      final: string[];
  }

  Per-stage ranked trace IDs for the hybrid retrieval pipeline (dense → sparse → merged → reranked → final). Populated by HybridRetriever so downstream consumers can compute per-stage retrieval-quality metrics (Recall@K, NDCG@K, MRR) and attribute losses to the stage that caused them. Absent for non-hybrid retrieval paths.

  • dense: string[]

  • sparse: string[]

  • merged: string[]

  • reranked: string[]

  • final: string[]

• Optional retrievedTypedTraces?: ScoredMemoryTrace[]

  Stage E: optional Hindsight typed-network output as canonical-shaped scored traces. When the manager is configured with typedNetwork and the variant supports retrieval-side activation ('full'), the manager delegates to a TypedNetworkRetriever which performs seed-finding (proper-noun + quoted-string entity extraction, case-insensitive intersection), spreading activation, and top-K ranking. Top-K results are surfaced as ScoredMemoryTrace[] for drop-in compatibility with the canonical retrieval pipeline (bank-prefixed content, namespaced IDs typed-network:<factId>, sourceType 'typed_network').

  Absent when typed-network is not configured. Empty when the retriever found no seed matches in the typed-network store.

  Phase 4.3 MVP: surfaced in diagnostics but NOT merged into the primary retrieved ranking. Phase 4.4 fusion lands when consumers wire the merged ranking. See 2026-04-26-hindsight-4network-observer-design.md.