[CLAUDE] Docs: S41 RAG audit response — exclude **/-anchor fix + retire stale _decision_log + AI_INFRA signal
All checks were successful
Deploy SOLUTION_ERP / build-deploy (push) Successful in 4m7s
All checks were successful
Deploy SOLUTION_ERP / build-deploy (push) Successful in 4m7s
- rag.json exclude_paths root-anchored -> **/-anchored (defeats gotcha #10: node_modules/** + docs/_archive/** were not matching nested paths) - _decision_log: retire stale "+321% / LIVE 11,922" -> real status (LIVE ~3080 ~= registry 3076, drift closed 2026-05-28) - New docs/governance/RAG-AUDIT-RESPONSE-2026-05-29.md: SE-side prep done + corrections (store_memory at-risk = 3 disk-backed broadcasts, NOT ~27) + re-bootstrap ask for AI_INFRA + post-bootstrap verify checklist Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
33
docs/governance/RAG-AUDIT-RESPONSE-2026-05-29.md
Normal file
33
docs/governance/RAG-AUDIT-RESPONSE-2026-05-29.md
Normal file
@ -0,0 +1,33 @@
|
||||
# 📤 SOLUTION_ERP → AI_INFRA — RAG Audit Response (2026-05-29, S41)
|
||||
|
||||
Re: AI_INFRA RAG audit 2026-05-29 (Qdrant LIVE verify). SE-side prep DONE; re-bootstrap = AI_INFRA op (charter v2). Persistent + corpus-backed record.
|
||||
|
||||
## ✅ SE-side DONE (this session)
|
||||
|
||||
1. **Exclude fix (`.claude/rag.json`)** — root-anchored → `**/`-anchored, defeats gotcha #10:
|
||||
- `node_modules/**` → `**/node_modules/**`
|
||||
- `docs/_archive/**` → `**/_archive/**` (also `bin`/`obj`/`.git` → `**/`-anchored for consistency)
|
||||
- JSON validated. Takes effect on next re-bootstrap.
|
||||
3. **`_decision_log` stale numbers retired** — `registry_drift_note` "+321% / LIVE 11,922" was pre-bootstrap STALE → rewritten to real status (LIVE ~3080 ≈ registry 3076, drift closed 2026-05-28). `anatomy_threshold_chosen` "11,922 mature" → "SE collection ~3080".
|
||||
|
||||
## 🔶 SE-side findings (corrections to audit estimates — verified on disk + Qdrant)
|
||||
|
||||
- **store_memory at-risk ≪ "~27".** True store_memory chunks (`heading_path="(manual)"`) = only the **3 S40 broadcasts**, ALL disk-backed (`docs/governance/BROADCAST-OUT-*.md` confirmed on disk). Replace-mode recreates them from corpus files → **NOT at-risk, no export-reinsert needed.** The "~27" appears to conflate with the 27 user-memory *feedback* entries — those are extra_corpus FILE-based; they need re-index to be **added**, not protected.
|
||||
- **node_modules junk confirmed:** `docs/_user-guide/node_modules/` = 30 `.md` files on disk (≈237 chunks plausible).
|
||||
- **_archive risk is now WORSE, not stable:** `docs/_archive/` now holds the 170KB+ pre-S40 STATUS/HANDOFF archives (created S40, after the 05-28 bootstrap). A re-bootstrap WITHOUT the exclude fix would index hundreds of archive chunks. Exclude fix must land before re-bootstrap.
|
||||
- **⚠️ Slug anomaly for AI_INFRA to confirm:** feedback chunks currently index under the OLD slug path `...projects\D--Dropbox-CONG-VIEC-SOLUTION\memory\` (missing `-SOLUTION-ERP`). Confirms the slug bug; replace-mode should wipe old-path chunks + re-add from the corrected `extra_corpus` path (rag.json:18, fixed S40).
|
||||
|
||||
## 🟢 ASK — AI_INFRA re-bootstrap (1 run, gathers everything)
|
||||
|
||||
`python AI_INFRA/claude-rag/bootstrap.py --project solution_erp` — picks up: (a) exclude fix → 0 node_modules + 0 _archive chunks; (b) corrected extra_corpus slug → 27 feedback entries indexed; (c) S38–S41 content (Proposal/WorkflowApps/consolidated docs).
|
||||
|
||||
Repeat of prior standing items (broadcast 2026-05-29): bootstrap.py corpus-path validation (warn on glob→0 files), verify auto_reindex hook actually fires (last_indexed lagged), search_code corpus gap (src/*.cs + fe/*.tsx not in corpus), registry sync.
|
||||
|
||||
## 🔍 SE post-bootstrap verify (after AI_INFRA confirms run)
|
||||
|
||||
1. node_modules chunks = 0 · _archive chunks = 0 (search a known junk term → expect miss)
|
||||
2. 27 feedback entries discoverable under corrected slug
|
||||
3. 3 broadcasts still present · chunk_count sane (no bloat)
|
||||
|
||||
## Stance (charter v2)
|
||||
SE = USER of infra. SE handled its own config declaration (corpus/exclude) + content; RAG mechanism (bootstrap/chunk/path-resolution) stays AI_INFRA. Conflict → anh pqhuy quyết.
|
||||
Reference in New Issue
Block a user