KB-BD8B

GPT Review — VRC Agent Prompt Approval with Amendments

4 min read Revision 1
vrciu-0vectorgpt-reviewdieu44agent-data

GPT Review — VRC Agent Prompt Approval with Amendments

Date: 2026-05-02 Scope: Review Opus VRC Runtime Investigation prompt + IU-0 vector direction Status: APPROVED WITH MANDATORY PATCHES

Verdict

The VRC prompt is directionally correct and should be dispatched to a runtime/code-capable Agent, but only after applying mandatory patches below.

Core architecture decision

  1. Legacy Agent Data Vector remains the production RAG route for all KB documents.
  2. IU Vector / IU Bridge must be a parallel track for information_unit, not a replacement.
  3. Every KB upload continues to be vectorized by the existing legacy pipeline.
  4. IU design must not alter collection, chunking, ingestion, trigger/listener, orphan audit, or re-embed behavior before VRC evidence + APR.
  5. Duplicate content between KB draft and IU final is expected to be very common and must be treated as a first-class VRC question.

Mandatory prompt patches

Patch 1 — No contradiction between read-only and report upload

The prompt currently says READ-ONLY / no mutation, while also asking the Agent to upload a report to KB. Clarify:

  • Infrastructure/data/vector/code investigation is read-only.
  • Report creation is the only allowed write, and only if the orchestration explicitly permits it.
  • If not permitted, Agent returns markdown report to User/GPT for upload.

Patch 2 — Do not call any endpoint that may cleanup/mutate

For /kb/audit-sync and similar endpoints:

  • Agent must inspect source first.
  • Agent may call only if confirmed read-only.
  • If endpoint triggers cleanup/delete/update/reconcile, do not call; report UNKNOWN and cite source.

Patch 3 — Discover DB/container names, do not assume

Commands using postgres, directus, or specific DB names are examples only. Agent must discover actual container/database/table before querying.

Patch 4 — Add VRC-11 Search retrieval behavior

In addition to VRC-10 duplicate content, Agent must inspect actual search/chat retrieval code:

  • Which collection(s) are queried?
  • Is payload filtering supported today?
  • Are filters exposed by API?
  • Does search merge/dedup results?
  • Does reranking exist?
  • What fields are returned to caller?
  • How top-K is constructed when two nearly identical vectors exist?

Patch 5 — Duplicate resolution must be evidence-based

VRC-10 must not just list options. It must recommend based on actual Qdrant/search code evidence:

  • Option A: separate collection for IU.
  • Option B: same collection with source_type/payload filter.
  • Option C: search both then dedup/rerank by canonical authority.
  • Option D: legacy-only fallback until IU pilot. The report must specify which option is safest for Phase 0, and why.

Patch 6 — Explicit “do not touch legacy invariants”

Agent must produce a final list of invariants that IU-0 cannot change without a separate APR and rollback plan.

Dispatch approval

After patches, dispatch VRC Agent Investigation before any IU-0 full draft. Do not upload IU-0. Do not open P44-6.

Next step after VRC report

  1. Read VRC report.
  2. Verify evidence quality.
  3. Decide IU vector architecture:
    • Legacy vector untouched.
    • IU vector/bridge parallel.
    • duplicate resolution strategy selected for pilot.
  4. Only then revise IU-0 OUTLINE-C and later full draft.