GPT Review — P3D Vector/Search Freshness Audit + Stop Pack 1

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

• knowledge/dev/laws/dieu44-trien-khai/reports/p3d-vector-search-freshness-audit-report.md
• Opus supplemental investigation on TAC/IU parallel systems

Verdict

STOP P3D Pack 1 continuation temporarily.

The production KB vector/search layer is a foundation dependency for the text-as-code build process. It must be stabilized before continuing further TAC↔IU work.

Main finding

The current observed search problem is not vector freshness and not missing vectors for the recent P3D documents.

Agent evidence:

• Five recent target documents exist in KB/PG.
• All five have vector_status=ready.
• All five have Qdrant points.
• All five are retrievable in top 20.
• Two exact-ish targets are not rank 1 because older semantically similar documents outrank them.

Confirmed root cause:

SEARCH_RANKING_NO_PATH_TITLE_BOOST_NOT_VECTOR_FRESHNESS

Confidence: high.

Is TAC technical debt causing this issue?

Direct cause: No

The current unreliable searchKnowledge behavior is in the legacy KB vector/search layer:

• Qdrant collection: production_documents.
• Legacy payload: document_id, content, metadata, chunk fields.
• Normal search path: Qdrant vector first.
• PG keyword fallback only runs when Qdrant has no hits or errors.
• No path/title/document_id boost in normal Qdrant path.

This exists independently of TAC/IU reconciliation.

Indirect future risk: Yes

Opus is correct that TAC and IU are currently parallel systems:

• TAC has 86 real law units and render proof.
• Native IU has gateway/edit/save machinery and pilot rows.
• Reconciliation is not done yet.

This does not explain today’s exact search-ranking failure, but it is a real architecture debt that will make future vector/search worse if ignored:

• duplicate KB/TAC/IU representations can create duplicate search results;
• legacy vector has no unit_id, canonical_address, or unit_version_id;
• IU vector is not implemented yet;
• without TAC↔IU reconciliation, future unit-level vector cannot be canonical.

Therefore:

• Fix legacy KB search ranking first.
• Then continue TAC↔IU reconciliation.
• Do not implement IU vector yet until canonical unit boundary is settled.

Production risk assessment

Current vector system is operating but not sufficiently reliable for production-grade agent workflow:

• Qdrant/Postgres/OpenAI are OK.
• Recent docs are vectorized.
• sync_status=warning still exists.
• Search ranking is noisy for exact path/title/document lookup.
• Health warning is mostly ratio-threshold + noisy ghost accounting, not missing recent P3D docs.

This is acceptable for ad-hoc semantic recall, but not acceptable as a production planning dependency where agents must reliably find exact current reports/prompts.

Required pause

Pause:

• Pack 1 implementation.
• Pack 1 DDL planning beyond read-only inventory.
• IU vector design/implementation.
• Any migration based on search assumptions.

Allowed:

• Read-only Pack 1 inventory if already queued.
• Vector/search fix design and review.
• Production-safe search behavior patch after review.

Next required pack

Create and execute a production-safe pack:

P3D_VECTOR_SEARCH_HYBRID_PATH_TITLE_BOOST_PACK

Goal:

• Make exact path/title/document_id queries deterministic or near-deterministic.
• Preserve legacy Qdrant collection and payload.
• Avoid reindex as first response.
• Avoid touching IU vector.
• Keep changes small, reversible, and tested.

Pack scope recommendation

Design first, then implementation only after GPT/User review.

Required capabilities:

1. Exact document_id / path slug boost.
2. metadata.title exact/substring boost.
3. Optional keyword overlap boost using title + path + body snippet.
4. Hybrid rerank over Qdrant results and/or PG keyword candidates.
5. Deterministic behavior for exact filenames such as p3d-pack1-readonly-inventory-prompt.md.
6. Search output should preserve original Qdrant score and include boost reason/debug fields if possible.
7. Audit health warning should not confuse empty/folder/short docs with real missing recent docs.

Hard boundaries

• No Qdrant collection replacement.
• No bulk reindex as first step.
• No auto-heal.
• No deleting/upserting points except if implementation requires normal search-code-only change; prefer no vector writes.
• No IU vector implementation.
• No TAC↔IU migration.
• No schema migration unless explicitly approved in a later pack.
• No production deploy without tests and rollback.

Status

vector_audit_review=PASS_ACCEPTED
root_cause=SEARCH_RANKING_NO_PATH_TITLE_BOOST_NOT_VECTOR_FRESHNESS
pack1_status=PAUSED_FOR_VECTOR_SEARCH_RELIABILITY
next_required_pack=P3D_VECTOR_SEARCH_HYBRID_PATH_TITLE_BOOST_PACK
implementation_allowed=false_until_prompt_review