Summary

• Adds per-strategy weights to RRF fusion, allowing tuning of how much each retrieval method (semantic, BM25, graph, temporal) influences the merged ranking
• Weights follow the hierarchical config model: server env vars → per-bank config → per-request override
• Default weights are all 1.0, preserving existing behavior exactly

API

New optional retrieval_weights field on the recall request:

{
  "query": "What happened at the meeting?",
  "retrieval_weights": {
    "semantic": 1.0,
    "bm25": 1.0,
    "graph": 2.0,
    "temporal": 1.0
  }
}

Keys: semantic, bm25, graph, temporal. Values are multipliers (1.0 = default, 2.0 = double influence, 0.0 = disabled). Omitted keys default to bank/server config.

Configuration

Server-level defaults via env vars:

• HINDSIGHT_API_RECALL_WEIGHT_SEMANTIC (default: 1.0)
• HINDSIGHT_API_RECALL_WEIGHT_BM25 (default: 1.0)
• HINDSIGHT_API_RECALL_WEIGHT_GRAPH (default: 1.0)
• HINDSIGHT_API_RECALL_WEIGHT_TEMPORAL (default: 1.0)

All four are in _CONFIGURABLE_FIELDS, so they can be overridden per-bank via the config API.

Changes

• fusion.py — reciprocal_rank_fusion() accepts optional weights dict; formula becomes w_i / (k + rank)
• config.py — 4 new hierarchical config fields with env var + default support
• http.py — RecallRequest.retrieval_weights field
• memory_engine.py — Merges config + request weights, passes to RRF
• interface.py — retrieval_weights param on abstract recall_async
• Control plane proxy + client types updated

Test plan

• 10 unit tests covering: unweighted default, None/1.0 equivalence, high weight boosting, zero weight disabling, partial weights, temporal support, ranking reversal, source rank preservation
• CI passes