Skip to content

KG-2.22 — Pack-Driven Retrieval Signals

Pillar: 2 — Epistemic Knowledge Graph · Status: live

What

Three declarative retrieval signals carried on the active Schema Pack and applied in the hybrid-retrieval hot path:

  • Recency decay — per node-type temporal boost measured against bi-temporal event_time (KG-2.11). Exponential or hyperbolic; always >= 1.0 so unknown dates are never penalised.
  • Source trust — per-source authority multiplier (e.g. peer_reviewed 1.3, arxiv 1.2, blog 0.7).
  • Autocut — trims the long tail at the largest relative score drop ("knee"), recall-safe via autocut_min_results.

These close the gap to gbrain's recency-decay.ts, source-boost.ts, and autocut.ts while remaining a strict no-op under the default core pack.

Why

Retrieval quality at scale needs more than vector similarity: fresh material should surface, trusted sources should outrank noise, and weak tails should be dropped rather than dilute the context window. Baking these into the pack means a domain can tune them without code changes.

How / Wiring

  • Declared on SchemaPack: recency_decay, source_trust, autocut_* (models/schema_pack.py), with helpers recency_spec_for() / trust_for().
  • Applied in HybridRetriever.retrieve_hybrid (knowledge_graph/retrieval/hybrid_retriever.py): _recency_boost(node, as_of) and _source_trust_boost(node) multiply _score in the post-fusion boost region; autocut() (retrieval/autocut.py) trims the ranked base set before graph traversal.
  • Entry point: graph_search(..., as_of=...) → engine.search_hybrid(as_of=...) → retrieve_hybrid. The as_of parameter enables "knowledge state as of date D".

Tests

tests/knowledge_graph/test_pack_retrieval_signals.py, tests/knowledge_graph/test_autocut.py.