KB-32AA

GPT Review — Vector Search Boost PASS, but Production Hardening Required

4 min read Revision 1
gpt-reviewvector-searchproduction-hardeningp3d2026-05-11

GPT Review — Vector Search Boost PASS, but Production Hardening Required

Date: 2026-05-11 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed:

  • knowledge/dev/laws/dieu44-trien-khai/reports/p3d-vector-search-hybrid-path-title-boost-implementation-report.md
  • knowledge/dev/laws/dieu44-trien-khai/reports/opus-verification-vector-search-boost-pass-2026-05-11.md

Verdict

Implementation PASS accepted, but not production-complete yet.

Search boost behavior is materially improved and tests pass. However, because this is production Agent Data search and future IU vector will reuse the same search/rerank mechanism, we must complete a hardening pack before resuming Pack 1.

Accepted as PASS

  • Root cause fixed at correct layer: app-layer rerank/path-title boost.
  • Qdrant unchanged: 4994 points before and after.
  • No PG schema mutation.
  • No reindex.
  • Health after patch: healthy.
  • T1 improved rank 5 → 1.
  • T2 improved rank 3 → 1.
  • T3 no regression.
  • T4 rank 2 → 2 is acceptable because baseline was rank 2; prompt wording was wrong, not implementation.
  • T5/T6 semantic regression tests pass.
  • Opus independent verification confirms hard cases rank well.

Critical production issue remaining

Agent report says compose uses baked image agent-data-local:latest, no build: and no source bind-mount. Patch was applied by:

host repo edit + docker cp into running container + docker compose restart

This means the patch is currently active, and host repo has commit ff2fc25, but the running container dependency path must be hardened:

  • If container is recreated from existing image, the docker cp patch may be lost.
  • If image is rebuilt from committed host repo, patch should be baked in, but this must be verified.
  • Current state is acceptable as a hotfix, not as final production deployment hygiene.

Additional non-blocking items now promoted to required hardening

Because User requested production-grade vector/search reliability before continuing:

  1. Deployment durability: ensure search boost survives container recreate/rebuild.
  2. Search regression canary: small repeatable test suite for T1–T6 + a few hard cases.
  3. Conservative recency tie-break: slight boost only when exact/path/title confidence is similar, not broad semantic displacement.
  4. Ghost/audit warning cleanup: classify empty/folder/short docs and empty-id edge cases so health warnings are meaningful.
  5. Unified search contract: document that legacy KB search and future IU vector search will share rerank/hybrid behavior; only chunk boundary differs for IU.
  6. Future IU vector compatibility: do not implement IU vector now, but design rerank signals so they can support canonical_address, unit_id, unit_version_id, content_hash when IU collection exists.

Decision

Keep Pack 1 paused. Do not resume TAC↔IU reconciliation until the Vector Search Reliability Hardening Pack is designed and either implemented or explicitly waived by User/GPT.

Required next pack

P3D_VECTOR_SEARCH_RELIABILITY_HARDENING_PACK

This pack should be high-level enough for Opus 4.6: one design + one implementation prompt, not many tiny prompts.

Status

vector_search_boost_behavior=PASS
production_durability=NOT_YET_PROVEN
pack1_status=PAUSED
next_required_pack=P3D_VECTOR_SEARCH_RELIABILITY_HARDENING_PACK
implementation_allowed=false_until_hardening_prompt_review
Back to Knowledge Hub knowledge/dev/laws/dieu44-trien-khai/reviews/gpt-review-vector-search-boost-pass-but-hardening-required-2026-05-11.md