KB-7E37

GPT Review — Vector Search Reliability Hardening Design + Prompt rev1

4 min read Revision 1
gpt-reviewvector-searchhardeningprompt-reviewp3d2026-05-11

GPT Review — Vector Search Reliability Hardening Design + Prompt rev1

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

  • knowledge/dev/laws/dieu44-trien-khai/design/p3d-vector-search-reliability-hardening-design.md
  • knowledge/dev/laws/dieu44-trien-khai/prompts/p3d-vector-search-reliability-hardening-implementation-prompt.md
  • knowledge/dev/laws/dieu44-trien-khai/reports/p3d-vector-search-reliability-hardening-design-report.md

Verdict

Design direction approved. Implementation prompt rev1 requires production-safety patch before dispatch.

The overall vision is correct: search/vector reliability must be hardened once, broadly enough to support both current KB search and future IU vector search. However, rev1 mixes several production-risk operations and must be tightened before execution.

Accepted

  1. Compose/build durability must be solved; docker cp hotfix is not enough.
  2. Canary/regression suite is required.
  3. Recency tie-break is acceptable if tiny and only applied after path/title boost.
  4. Audit warning noise must be addressed so sync_status is meaningful.
  5. Unified search contract for legacy KB + future IU vector is required.
  6. IU vector remains deferred; only compatibility contract/shape is prepared.

Required corrections

  1. Do not allow PG update for test_empty.md.tmpl in this pack. Audit cleanup should be code-level classification/exclusion, not row mutation. No DB writes.

  2. Compose/build change must be guarded. Add build: only after confirming Dockerfile and compose behavior. Back up compose. Rebuild only agent-data. Do not touch Qdrant.

  3. Force-recreate must be limited. Only agent-data may be recreated. Do not recreate the full stack.

  4. Canary parser must be inspected/adapted to actual MCP/API response. The template parser may be wrong. Agent must first capture actual response shape.

  5. Recency tie-break must not depend on metadata fields that may be absent. If updated_at/created_at is not in Qdrant payload, record warning and skip recency rather than inventing data or changing payloads.

  6. Audit cleanup must not change data/vector state. It may change health/audit classification code to exclude empty/folder/short/non-vectorizable docs, but no auto-heal/reindex/delete/upsert.

  7. Unified IU contract should be documented and code-shaped minimally. Add optional collection_name/source parameter only if low-risk; no IU collection, no IU payload, no new Qdrant index.

  8. Durability proof must include recreate test and canary after recreate. Search boost must survive docker compose up -d --force-recreate agent-data.

Required next action

GPT will patch the implementation prompt to rev2 directly. Opus should not write another micro-review unless rev2 introduces a substantive issue.

Status

hardening_design=APPROVED
implementation_prompt_rev1=NOT_APPROVED_FOR_DISPATCH
required_next=GPT_PATCH_REV2_THEN_AGENT_DISPATCH
pack1_status=PAUSED
Back to Knowledge Hub knowledge/dev/laws/dieu44-trien-khai/reviews/gpt-review-vector-search-hardening-design-and-prompt-rev1-2026-05-11.md