Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Noise-Aware Graph + Hub Dampening (Discern)

File: src/search/graph_ppr.rs (type_base_weight, dampen_hubs)

The problem

The live knowledge graph was ~94% taxonomy noise: tagged_with edges (note → tag noun) dwarfed the ~134 semantic edges, and degree-73/101/150 mega-hubs let PPR mass wash out across tag clouds. Unweighted PPR on such a graph returns noise. And a query that looked “too vague” to answer (abstention) never got a graph chance at all.

The references

  • GAAMA (arXiv:2603.27910) — hub dampening w_ij · min(1, θ/deg(i)) tames mega-hubs; edge-type weights separate taxonomy from semantics.
  • MemORAI (arXiv:2605.01386) — static-type weighting.
  • “Use Graph When It Needs” (arXiv:2602.03578) — complexity-gated activation: engage the graph leg precisely when the estimator says it helps.

The implementation (v1.12.0 “Discern”)

  1. Edge-type weights: type_base_weight — tagged_with/alias_of → 0.1, all other relation types → 1.0. The pair-aggregation SQL groups by relation_type, scales each group by its type weight, then sums per pair.
  2. Hub dampening: SparseGraph::dampen_hubs(θ) with HUB_DAMPING_THETA = 50 — GAAMA’s per-source min(1, θ/deg(i)), applied to the reachable-bounded graph before PPR. Per-source asymmetry is intentional (matches the reference). Determinism hardened by sorting edge rows.
  3. Complexity-gated rescue: should_attempt_graph_rescue fires a bounded graph-augmented pass only when the estimator says ClarifyQuery, the graph leg isn’t already on, and BRAIN_GRAPH_RESCUE_ENABLED (default true). abstention_decision returns low_confidence only when ClarifyQuery AND the final hit list is empty — a successful rescue returns its hits with decision: "ok", strictly additive, no behavior regression when the kill switch is off.

Measured ceiling

  • θ=50 and the 0.1 type weight are corpus-calibrated constants, not learned (deterministic + auditable by design).
  • The rescue fires only on the would-be-abstention path; a query with no KG structure (no entity match → no seeds) still abstains.
  • Type weights are static (no query conditioning); concept nodes (GAAMA), query-conditioned weights (MemORAI), and noun-phrase seeding remain future options. The tag cloud is structural — re-created on every re-ingest.

Pinned by a regression test that temporarily reverting to the v1.11 arithmetic fails — the mechanism is proven, not asserted.