Phase 5 — Goals

Track A — Hybrid Retrieval (Dense + Sparse + Reranking)

Goal 5.A: Combine dense and sparse retrieval with cross-encoder reranking

Milestones:

• 5.A.1 — Sparse (BM25) retrieval channel: find_lexical() using ts_rank_cd (cover density), parallel sibling to find_similar. GIN index verified at 100K+ row scale.
• 5.A.2 — Reciprocal Rank Fusion combiner: reciprocal_rank_fusion() combining dense and sparse result lists. Memories appearing in both lists near the top dominate.
• 5.A.3 — Cross-encoder reranking: BAAI/bge-reranker-v2-m3 via TEI /rerank endpoint. Only runs on top-50 RRF-fused candidates. Graceful degradation to RRF-only on TEI unavailability.

Acceptance criteria:

• Exact-phrase query ranks exact-match memory above semantically-similar-but-wrong one
• Hybrid beats pure-dense baseline on gold-set v2 lexical queries
• Reranker p95 latency fits within 40ms context-assembly deadline as explicit sub-budget
• Degradation to RRF-only on TEI timeout verified

---

Track B — Graph-Aware Retrieval

Goal 5.B: Traverse the memory relationship graph at query time for lineage and contradiction awareness

Milestones:

• 5.B.1 — Graph traversal query layer: Recursive CTE in Postgres, capped at 5 hops. org_id filter at every level of the recursion. Cycle detection via depth cap.
• 5.B.2 — Graph-boosted retrieval scoring: Superseded memories never outrank their successor. Unresolved contradictions in top-K: only higher-scored one packed into context.
• 5.B.3 — MCP get_memory_lineage tool: Returns full supersession/contradiction/specialization chain. Same auth/rate-limit path as existing MCP tools.

Acceptance criteria:

• org_id filter present at every CTE level (cross-org traversal-leak test)
• Contradiction-pair fixture: only higher composite-score memory survives into packed context
• ISO-MCP-LINEAGE-01 cross-tenant test passes

---

Track C — Evaluation Harness Upgrade

Goal 5.C: Multi-mode retrieval benchmark with online shadow evaluation

Milestones:

• 5.C.1 — Hybrid-aware gold set v2: ~500 memories, 150 labeled queries. Lexical-specific and graph-traversal-specific subsets. Two human labelers, >= 0.8 Cohen's kappa.
• 5.C.2 — Multi-mode benchmark runner: Dense-only, hybrid, hybrid+rerank, full-graph configs run side by side against the same gold set.
• 5.C.3 — Production shadow evaluation: 1% of real retrievals run through both live and candidate configs. Divergence logged to ClickHouse. Weekly automated report on divergence rate.

Acceptance criteria:

• At least 30 queries where dense-only is measurably worse than hybrid (proves RRF value)
• At least 20 queries requiring graph traversal (proves Track B value)
• CI fails on regression against committed baseline (3 point tolerance)
• Shadow path adds zero latency to real requests

---

Track D — Exit Gate

Goal 5.D: Verify advanced retrieval quality and performance at phase boundary

Acceptance criteria:

• All exit criteria from the phase index met
• Phase 5 gap register published with zero open P0s
• ADR index updated