Phase 5 — Advanced Retrieval & Graph Memory

Status: Planned
Depends on: Phase 4.5 — Intelligence Layer Next: Future roadmap work (org-wide production hardening remains deferred)

Purpose

Phase 3 established a flat vector store with a memory_relationships graph table that is not yet traversed at query time. Phase 5 is where retrieval quality moves beyond "one embedding, one ANN search, one score" toward hybrid search, reranking, and graph-aware packing:

• Hybrid search: Dense (HNSW) + Sparse (BM25 via Postgres GIN) combined with Reciprocal Rank Fusion
• Cross-encoder reranking: BAAI/bge-reranker-v2-m3 via TEI for top-50 candidate refinement
• Graph-aware retrieval: Multi-hop traversal of memory_relationships for lineage/contradiction chains
• Production shadow evaluation: 1% traffic sampling for online retrieval-quality monitoring

Suggested paths and libraries below are planning defaults. Implementation may revisit them after measurement, further research, or operational constraints.

Milestones

What Phase 5 is not

• Org-wide observability stack / chaos engineering / release governance (future work)
• Replacing the Phase 3 dense retrieval path as the only available mode (dense-only remains a baseline)
• A separate graph database requirement (Postgres recursive CTEs are the starting preference)

Entry criteria

• Phase 4.5 exit criteria met
• Phase 3 memory_relationships populated and Phase 3.5 context assembly live

Exit criteria

1. Hybrid retrieval (BM25 + RRF) live behind feature flag, default-on, beating dense-only on gold-set lexical queries
2. Cross-encoder reranker meets latency budget without breaking context assembly p95 <= 50ms
3. Graph traversal live via API and MCP get_memory_lineage tool with cross-tenant isolation test passing
4. Gold-set v2 committed with inter-annotator agreement >= 0.8 kappa
5. Multi-mode benchmark running in CI on every retrieval-touching PR
6. Shadow-evaluation pipeline live with divergence dashboard
7. All Phase 1–4.5 regression suites still green
8. make e2e-smoke-p5 exits 0
9. Phase 5 gap register published with zero open P0s
10. Org-wide production hardening remains explicitly deferred as future roadmap scope

Execution order

Track A: 5.A.1 → 5.A.2 → 5.A.3
Track B: 5.B.1 → 5.B.2 → 5.B.3 (can start after Phase 4.5; pairs with Track A for eval)
Track C: 5.C.1 → 5.C.2 → 5.C.3 (depends on A/B surfaces existing)
Track D: 5.D.1 → 5.D.2 (after A–C verification)

Documents

• tracks.mdx — Track purpose and milestone grouping
• goals.mdx — Goal overview by track
• risks.mdx — Risks and mitigations
• milestones/ — PR-sized work units

Goal overview