文件内容
configs/retrieval/defaults.yaml
top_k_reply_pairs: 8
top_k_documents: 3
top_k_chunks: 3
recency_boost_days: 60
recency_boost_weight: 0.2
account_boost_weight: 0.15
# Semantic search settings
semantic_weight: 0.4
semantic_min_coverage: 0.01
# Sender-aware retrieval boosts
sender_type_boost: 0.15
sender_domain_boost: 0.10
# Per-sender-type boost multipliers (applied to match scores)
# Cross-encoder reranking (requires sentence-transformers)
reranker_enabled: false
# Boost applied when a reply-pair's subject literally matches query tokens.
# Historically hardcoded at construction time; surfaced for autoresearch
# tuning. Multiplicative on the combined score.
subject_match_boost: 0.2
topic_match_boost: 0.15
# Per-token contribution to `_field_match_bonus` (title/subject literal-match
# bonus on top of BM25). Historically the literal `0.25` inside the helper.
field_match_bonus_per_token: 0.25
# BM25 lexical-score saturation. The historical hardcoded `min(rank*2, 10)`
# is preserved here; surfaced so the autoresearch loop can A/B-tune them
# against the golden-eval gate. scale ramps how fast lexical_score grows per
# unit of FTS5 rank; cap is the ceiling above which all matches tie.
lexical_scale: 2.0
lexical_cap: 10.0
sender_type_boost_map:
external_client: 1.3
personal: 0.8
automated: 0.3
internal: 1.5