Tracks
Phase 5 tracks for hybrid retrieval, graph-aware retrieval, evaluation, and exit gate.
Phase 5 — Tracks
Phase 5 is organized into four tracks. Paths and libraries named below are planning defaults and may change after measurement or further investigation.
Track A — Hybrid Retrieval
Purpose: Close the exact-match gap of pure dense retrieval by adding a parallel BM25 channel, fusing dense and sparse rankings with Reciprocal Rank Fusion, then optionally reranking a bounded candidate set with a cross-encoder.
Preferred sequencing within track: 5.A.1 → 5.A.2 → 5.A.3
Related tracks: Track C evaluates whether hybrid and reranked modes beat dense-only.
Milestones:
Track B — Graph-Aware Retrieval
Purpose: Use the existing memory_relationships table at query time for lineage, supersession,
and contradiction awareness — starting with Postgres recursive CTEs rather than a separate graph DB.
Preferred sequencing within track: 5.B.1 → 5.B.2 → 5.B.3
Related tracks: Track A produces the candidate set; Track C includes graph-specific eval queries.
Milestones:
- 5.B.1 — Graph Traversal Query Layer
- 5.B.2 — Graph-Boosted Scoring
- 5.B.3 — MCP get_memory_lineage Tool
Track C — Evaluation Harness Upgrade
Purpose: Keep retrieval quality measurable as hybrid and graph features land — gold-set v2, side-by-side mode comparison, and optional production shadow sampling.
Preferred sequencing within track: 5.C.1 → 5.C.2 → 5.C.3 (5.C.3 can start after 5.C.2 once live configs exist)
Related tracks: Depends on Tracks A and B being far enough along to exercise real configs.
Milestones:
- 5.C.1 — Hybrid-Aware Gold Set v2
- 5.C.2 — Multi-Mode Benchmark Runner
- 5.C.3 — Production Shadow Evaluation
Track D — Exit Gate
Purpose: Sign off retrieval quality and publish the Phase 5 gap register before declaring the phase complete.
Preferred sequencing within track: 5.D.1 → 5.D.2
Milestones:
Last updated on