Module scoring 

Composite memory scoring — ported from the archived internal service (v1 archive).

Provides a fully-tunable scoring pipeline:

1. ScoringConfig — all knobs, all pub, all serde-friendly for agent sweeps.
2. calculate_score — multiplicative formula: w_rel × relevance × (1 + w_temp × recency) × (1 + w_imp × salience).
3. ScoreAdjustment — declarative conditional rules applied after the base formula.
4. normalize_rrf_scores / normalize_rank_fusion_scores — RRF and raw-cosine normalization.
5. normalize_min_score — dual-scale input (0.0–1.0 fraction or 0–100 integer).
6. is_meaningful_query — noise gate before embedding compute.
7. contains_cjk — CJK routing decision.
8. needs_multilingual — broad multilingual routing gate (non-ASCII alphabetic script).

Structs

CandidateContext

Context passed to condition evaluation — properties of the candidate + query.

ScoreAdjustment

A conditional score adjustment: if condition matches, apply operation.

ScoreInput

Input data for calculate_score — groups per-candidate fields to stay within clippy’s 7-argument limit.

ScoringConfig

Complete, tunable scoring configuration for the memory recall pipeline. DoS caps: max_recall_candidates≤500, default_token_budget≤16000, default_recall_limit≤200.

ScoringWeights

Weights for the combined memory score: score = w_rel × relevance × (1 + w_temp × recency) × (1 + w_imp × salience).

Enums

AdjustmentCondition

A condition that determines whether a score adjustment applies to a candidate.

AdjustmentOp

The operation to apply when a condition matches.

Constants

ENTITY_POSTERIOR_CLAMP_MAX

Upper bound of the per-entity posterior multiplier — the term can never move a score up by more than 15%.

ENTITY_POSTERIOR_CLAMP_MIN

Lower bound of the per-entity posterior multiplier — the term can never move a score down by more than 15%.

ENTITY_POSTERIOR_WEIGHT

Default weight for the per-entity posterior term. Chosen so the clamp bounds below are exactly reachable at posterior means of 0.0 and 1.0 (1 + 0.3 * (0.0 - 0.5) = 0.85, 1 + 0.3 * (1.0 - 0.5) = 1.15).

MAX_AUTO_ENTITY_NAMES

Maximum number of auto-extracted entity-name candidates per query. EntityMatch (see default_adjustments above) applies its multiplier to any candidate memory that contains any one of the extracted names — this constant does not limit how many memories can receive the boost (one generic name can still match many memories). It only bounds the extracted-name list length and, transitively, the number of contains_at_word_boundary scans EntityMatch performs per note.

MAX_RECALL_CANDIDATES

Maximum candidates a caller may request (server-side cap).

MAX_RECALL_LIMIT

Maximum result limit a caller may request (server-side cap).

MAX_TOKEN_BUDGET

Maximum token budget a caller may request (server-side cap).

Functions

calculate_score

Composite score for a single memory candidate.

contains_cjk

Returns true when >15% of the query’s characters are CJK.

default_adjustments

Default score adjustments: episodic recency bonus, semantic age penalty, entity boost.

entity_posterior_term

clamp(1 + w_ent * (entity_posterior_mean - 0.5), 0.85, 1.15).

extract_entity_candidates

Auto-extract entity-name candidates from a recall query when the caller does not supply entity_names explicitly.

is_cjk_char

Returns true if c is a CJK character (Unified, Extension A/B, Hiragana, Katakana, Hangul).

is_meaningful_query

Returns true if the query has enough semantic content for meaningful recall.

needs_multilingual

Returns true when text should be routed to the multilingual embedding model for dense retrieval.

normalize_min_score

Normalize min_score: 0–1 passes through, 1–100 divides by 100, others return Err.

normalize_rank_fusion_scores

Normalize raw-cosine or BM25 scores (single-source) into a calibrated relevance band.

normalize_rrf_scores

Normalize RRF-fused scores (dual-source) into a calibrated relevance band.