db(p11): rag_query_log + learning_episodes — RAG 自主學習迴圈基礎
All checks were successful
CD Pipeline / deploy (push) Successful in 3m30s
All checks were successful
CD Pipeline / deploy (push) Successful in 3m30s
Operation Ollama-First v5.0 / Phase 11 RAG + 自主學習 migrations/027 — rag_query_log(每次 RAG 查詢的 audit log) - query_text 4KB CHECK + 90 天保留 - VECTOR(1024) bge-m3 embedding (與 ai_insights 一致簽名) - ivfflat lists=100 索引 - saved_call 欄位追蹤「成功攔截 LLM 呼叫」次數 - feedback_score 1-5(NULL=未反饋) - 6 條 CHECK 含 chk_rag_saved_consistent migrations/028 — learning_episodes(蒸餾池 → ai_insights 前哨) - 8 狀態機:pending/approved/awaiting_review/rejected_*4/expired - weight 0-1(>=0.8 觸發 PromotionGate Stage 4 人工驗收) - 9 條 CHECK 含 chk_le_approved_consistent / chk_le_review_consistent - partial index idx_le_status WHERE in (pending, awaiting_review) - distilled_text 16KB 上限 docs/phase11_db_design — 設計文檔 - 6 大決策(兩表分離 / ivfflat / partial index / 軟連結 / 90天保留 / 應用層白名單) - 6 大風險評估(R1 PII / R2 蒸餾失誤 / R3 ivfflat 退化 / R4 dangling FK / R5/R6 trade-off) - Phase 11 上線後驗收 SQL(EXPLAIN ANALYZE) PromotionGate 4 階段(v5.0 護欄 #1, ADR-033): Stage 1: quality_score >= 0.7 Stage 2: 無幻覺檢測(規則引擎,零 LLM) Stage 3: 與既有 insight 相似度 < 0.95 Stage 4: weight >= 0.8 必經 Telegram 👍/👎(24h 無回應 → expired) A4 fullstack-engineer 同時在寫 services/rag_service.py + learning_pipeline.py, service 完成後一起部署啟用。 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
207
docs/phase11_db_design_20260503.md
Normal file
207
docs/phase11_db_design_20260503.md
Normal file
@@ -0,0 +1,207 @@
|
||||
# Phase 11 DB 設計:RAG + 自主學習迴圈
|
||||
|
||||
- **戰役**: Operation Ollama-First v5.0 — Phase 11
|
||||
- **作者**: A3 db-expert
|
||||
- **日期**: 2026-05-03 台北
|
||||
- **migration**: `migrations/027_create_rag_query_log.sql`、`migrations/028_create_learning_episodes.sql`
|
||||
- **對應 ADR**: ADR-029(Hermes-First)、ADR-002(pgvector 唯一向量庫)、ADR-007(pgvector 啟用)
|
||||
- **前置 migration**: 024(ai_calls)、025(mcp_calls + ai_call_budgets)、026(embedding_signature)
|
||||
|
||||
---
|
||||
|
||||
## 1. 為何分兩表(rag_query_log vs learning_episodes)
|
||||
|
||||
兩個表責任完全不同,混表會讓**讀寫模式衝突**且**保留週期混淆**。
|
||||
|
||||
| 維度 | rag_query_log | learning_episodes |
|
||||
|---|---|---|
|
||||
| **角色** | RAG 召回的 audit log | 知識庫前哨(蒸餾池) |
|
||||
| **資料方向** | 從用戶/呼叫者「進來」 | 給 ai_insights「出去」 |
|
||||
| **生命週期** | 90 天滾動刪除 | 長期(approved/rejected 走冷儲檔) |
|
||||
| **寫入頻率** | 高(每次 RAG 召回都寫) | 中(過 quality 才寫) |
|
||||
| **PII 風險** | 高(query_text = 用戶問題) | 低(distilled 已蒸餾) |
|
||||
| **典型查詢** | 「過去 24h 命中率」「caller 分布」 | 「待人工驗收清單」「Stage 3 dedup query」 |
|
||||
| **是否進 RAG 召回語料** | 否(只是 log) | 否(只有晉升 ai_insights 後才進) |
|
||||
|
||||
**反證**:若合表,會出現
|
||||
- query_text PII 與蒸餾文本同表→ 90 天保留無法分別套用
|
||||
- 高頻寫入 audit log 與低頻寫入蒸餾池共享 ivfflat 索引 → vacuum / REINDEX 衝突
|
||||
- promotion_status 對 audit log 無意義,但要忍受 NULL
|
||||
|
||||
故維持分表。
|
||||
|
||||
---
|
||||
|
||||
## 2. ivfflat lists=100 計算依據
|
||||
|
||||
pgvector 官方建議:
|
||||
- `lists ≈ rows / 1000`(rows < 1M)
|
||||
- `lists ≈ sqrt(rows)`(rows ≥ 1M)
|
||||
|
||||
**rag_query_log 量推估**:
|
||||
- 假設 Phase 11 上線後每日 RAG 召回 5,000 次(hermes_qa + openclaw_qa + 內部 caller)
|
||||
- 90 天保留 → 穩態約 **450k 行**
|
||||
- `lists ≈ 450 / 1`,但太小(<10)會退化成全掃;取下限 100
|
||||
- 等流量上升到 1M 行時(約 200 天後若日 5k → 不會到 1M),再 `REINDEX ... WITH (lists=1000)`
|
||||
|
||||
**learning_episodes 量推估**:
|
||||
- 假設每日蒸餾 200 筆(rejected ~70%、approved ~30%)→ 全保留
|
||||
- 一年約 73k 行;2 年約 146k 行
|
||||
- `lists=100` 在 1M 以下都合理
|
||||
|
||||
**重訓 SOP**(寫入 ADR-029 後續維運章節):
|
||||
```sql
|
||||
-- 每月由 scheduler 檢查,若 EXPLAIN cost / actual_time 退化 5x,重訓
|
||||
REINDEX INDEX CONCURRENTLY idx_rag_query_log_embedding;
|
||||
REINDEX INDEX CONCURRENTLY idx_le_embedding;
|
||||
```
|
||||
|
||||
**為何不用 HNSW(009 ai_insights 用 HNSW)**:
|
||||
- HNSW 寫入比 ivfflat 慢 5-10x(高頻寫入的 rag_query_log 不適合)
|
||||
- HNSW 不需訓練,但**索引大小**約為 ivfflat 的 2-4×
|
||||
- ai_insights 是「讀多寫少」(KM 沉澱)—— HNSW 合理
|
||||
- rag_query_log / learning_episodes 是「寫多讀中」—— ivfflat 合理
|
||||
|
||||
---
|
||||
|
||||
## 3. promotion_status 狀態機
|
||||
|
||||
```
|
||||
┌─────────────┐
|
||||
│ pending │ (初始)
|
||||
└──────┬──────┘
|
||||
│
|
||||
┌────────────┴───────────────┐
|
||||
│ │
|
||||
Stage 1: quality<0.7 Stage 2: 規則檢測幻覺
|
||||
│ │
|
||||
▼ ▼
|
||||
┌──────────────────┐ ┌───────────────────────┐
|
||||
│ rejected_quality │ │rejected_hallucination │
|
||||
└──────────────────┘ └───────────────────────┘
|
||||
|
||||
Stage 1+2 通過:
|
||||
│
|
||||
▼
|
||||
Stage 3: 與既有 insight cosine>0.95
|
||||
│
|
||||
▼
|
||||
┌────────────────────┐
|
||||
│ rejected_duplicate │ (若太相似)
|
||||
└────────────────────┘
|
||||
|
||||
Stage 3 通過:
|
||||
│
|
||||
┌────────┴────────────┐
|
||||
│ │
|
||||
weight<0.8 weight>=0.8
|
||||
│ │
|
||||
▼ ▼
|
||||
┌──────────┐ ┌──────────────────┐
|
||||
│ approved │ │ awaiting_review │ ← Telegram 推播
|
||||
└──────────┘ └────────┬─────────┘
|
||||
│ │
|
||||
│ ┌──────────┼─────────────┐
|
||||
│ │ │ │
|
||||
│ 人工 👍 人工 👎 24h 無反饋
|
||||
│ │ │ │
|
||||
│ ▼ ▼ ▼
|
||||
│ ┌──────────┐ ┌──────────────┐ ┌──────────┐
|
||||
│ │ approved │ │rejected_human│ │ expired │── 降 weight=0.5 重走 Stage 4a
|
||||
│ └──────────┘ └──────────────┘ └──────────┘
|
||||
│ │
|
||||
▼ ▼
|
||||
寫 ai_insights → insight_id 回填
|
||||
```
|
||||
|
||||
**關鍵 invariants(已用 CHECK 強制)**:
|
||||
1. `approved ⇔ insight_id IS NOT NULL`(chk_le_approved_consistent)
|
||||
2. `rejected_* ⇒ rejected_reason IS NOT NULL`(chk_le_rejected_reason)
|
||||
3. `human_approver IS NOT NULL ⇒ reviewed_at IS NOT NULL`(chk_le_review_consistent)
|
||||
|
||||
---
|
||||
|
||||
## 4. 90 天保留策略
|
||||
|
||||
| 表 | 保留 | 工具 | 預計排程 |
|
||||
|---|---|---|---|
|
||||
| `rag_query_log` | 90 天 | scheduler `DELETE WHERE queried_at < NOW() - INTERVAL '90 days'` | 03:30 daily |
|
||||
| `learning_episodes` (pending/awaiting_review) | 永久(直到狀態變化) | — | — |
|
||||
| `learning_episodes` (approved) | 永久(蒸餾溯源) | — | — |
|
||||
| `learning_episodes` (rejected_*/expired) | 180 天後可冷儲檔 | 後續 ADR 定 | monthly |
|
||||
| `ai_calls` | 90 天 | (已存在 migration 024 註解) | 03:00 daily |
|
||||
| `mcp_calls` | 90 天 | 同上 | 03:15 daily |
|
||||
|
||||
**為何 rag_query_log 與 ai_calls 同 90 天**:兩者透過 `request_id` 串鏈;若不同保留期會出現「ai_calls 已刪、rag_query_log 留著」的孤兒,反查 trace 會斷。
|
||||
|
||||
**learning_episodes 不限期保留的依據**:蒸餾池是「為什麼這條 insight 進了 KM」的證據鏈。`rejected_*` 也保留是為了**防止同類錯誤被反覆生成**(PromotionGate Stage 3 dedup 可參考歷史 rejected)。
|
||||
|
||||
---
|
||||
|
||||
## 5. 風險評估
|
||||
|
||||
### R1(HIGH)—— query_text PII 落地
|
||||
|
||||
- **風險**:`rag_query_log.query_text` 是用戶原始輸入,可能含人名/手機/訂單號
|
||||
- **緩解**:
|
||||
1. CHECK `octet_length <= 4096` 限長度
|
||||
2. 90 天滾動刪除
|
||||
3. 應用層在寫入前對「明顯 PII pattern」做 redact(如 `\d{10}` 手機)
|
||||
4. `learning_episodes.distilled_text` 必須是「蒸餾後」文本,**禁止**直接複製 query_text
|
||||
- **未解殘留風險**:90 天內 DBA query 仍可看到原始問題;建議搭配 PostgreSQL row-level audit log 追蹤誰查過
|
||||
|
||||
### R2(HIGH)—— 蒸餾失誤污染 RAG
|
||||
|
||||
- **風險**:低品質 `learning_episodes` 過閘晉升 `ai_insights` → RAG 召回幻覺擴散
|
||||
- **緩解**:
|
||||
1. PromotionGate 4 階段(quality / hallucination / duplicate / human)
|
||||
2. `weight>=0.8` 強制人工驗收(chk_le_review_consistent)
|
||||
3. `rejected_*` 必填 rejected_reason(chk_le_rejected_reason),事後可審計
|
||||
|
||||
### R3(MEDIUM)—— ivfflat 索引膨脹 / 退化
|
||||
|
||||
- **風險**:高頻寫入 + 不重訓 → recall 退化
|
||||
- **緩解**:
|
||||
1. partial index `WHERE query_embedding IS NOT NULL` 縮體積
|
||||
2. monthly REINDEX CONCURRENTLY(見上 §2 SOP)
|
||||
3. EXPLAIN ANALYZE alarm(cost > baseline 5x 時告警)
|
||||
|
||||
### R4(MEDIUM)—— ai_insights 軟連結 dangling
|
||||
|
||||
- **風險**:`learning_episodes.insight_id` 無 FK,若 ai_insights archive,蒸餾池會留 dangling pointer
|
||||
- **緩解**:
|
||||
1. archive 時保留 ai_insights 主鍵(採 status='archived' soft delete,而非 DELETE)
|
||||
2. 應用層 join 用 LEFT JOIN,dangling 顯示為 "已歸檔"
|
||||
|
||||
### R5(LOW)—— used_results BIGINT[] 反正規化
|
||||
|
||||
- **風險**:`rag_query_log.used_results` 用陣列存命中 ai_insights.id,違反正規化
|
||||
- **緩解理由**:
|
||||
1. 召回每筆平均 3-5 個 id,若拆 join table 會 5x 寫入放大
|
||||
2. 反向查詢「某 insight 被多少 RAG 命中」是低頻分析,可用 `WHERE id = ANY(used_results)` 或 GIN 索引補(V2 再加)
|
||||
- **接受該風險**
|
||||
|
||||
### R6(LOW)—— caller 白名單未在 DB 強制
|
||||
|
||||
- **風險**:應用層可能寫入未知 caller,污染統計
|
||||
- **緩解**:
|
||||
1. ai_calls 已有 caller 白名單註釋,logger 統一強制
|
||||
2. 本表加 CHECK 會與 ai_calls 雙寫漂移;改由 application layer 單一真理源
|
||||
- **接受該風險**
|
||||
|
||||
---
|
||||
|
||||
## 6. 驗收清單(給 critic)
|
||||
|
||||
- [x] 027 / 028 連續編號,未跳號
|
||||
- [x] BIGSERIAL 主鍵對齊 024/025
|
||||
- [x] CHECK 風格對齊 critic-A11(白名單 + size + range)
|
||||
- [x] partial index 對 sparse 欄位(request_id / insight_id / status)
|
||||
- [x] ivfflat lists=100 + cosine + 1024 維對齊 bge-m3
|
||||
- [x] GRANT 權限對齊(momo + sequence)
|
||||
- [x] 不在 migration 內 CONCURRENTLY(無既存大表)
|
||||
- [x] 回滾腳本附在 migration 頂部註解
|
||||
- [x] 與 ai_calls/mcp_calls 透過 request_id 串鏈
|
||||
- [x] PII 護欄(query_text 4KB / distilled 16KB)
|
||||
- [x] 狀態機 invariant 用 CHECK 鎖死
|
||||
- [x] 不自動 commit / 不自動 apply
|
||||
126
migrations/027_create_rag_query_log.sql
Normal file
126
migrations/027_create_rag_query_log.sql
Normal file
@@ -0,0 +1,126 @@
|
||||
-- =============================================================================
|
||||
-- Migration 027: rag_query_log — RAG 查詢遙測 (audit log)
|
||||
-- Operation Ollama-First v5.0 — Phase 11
|
||||
-- 日期: 2026-05-03 台北
|
||||
-- 對應戰役: ADR-029(Hermes-First 雙塔)+ Phase 11 RAG 自主學習迴圈
|
||||
-- =============================================================================
|
||||
-- 說明:
|
||||
-- 每次 RAG 召回(hermes_qa / openclaw_qa / etc.)寫一筆,append-only。
|
||||
-- 核心指標:
|
||||
-- - hit_count : top_k 召回實際命中數(threshold 過濾後)
|
||||
-- - saved_call : 命中且最終未升級到 LLM => 真實節省成本
|
||||
-- - feedback_score : Telegram 👍/👎 後填回(NULL = 尚未反饋)
|
||||
-- 與 ai_calls / mcp_calls 透過 request_id 串鏈,跨表 trace 同一邏輯請求。
|
||||
--
|
||||
-- 設計決策:
|
||||
-- 1. embedding 與 ai_insights 同維度 1024 (bge-m3),可跨表計 cosine
|
||||
-- 2. ivfflat lists=100 對齊既有風格;資料量達 1M 後依 sqrt(N) 重建
|
||||
-- (009 ai_insights 用 HNSW,但本表寫入頻繁 + 不需即時最近鄰 query
|
||||
-- 採 ivfflat 寫入便宜,weekly 排程 REINDEX 即可,詳見 design doc)
|
||||
-- 3. used_results BIGINT[] 紀錄命中的 ai_insights.id,方便事後召回率分析
|
||||
-- (不加 FK;ai_insights 可能 archive,避免 cascade)
|
||||
-- 4. query_text 限 4KB;query_text 可能含 PII(用戶問題)—
|
||||
-- 90 天保留 + 後續 PromotionGate 過濾後才允許進 ai_insights
|
||||
-- 5. caller 與 ai_calls.caller 共白名單(不重複定義 CHECK,避免雙寫漂移;
|
||||
-- 由 application logger 端強制;DB 端僅檢長度)
|
||||
--
|
||||
-- 回滾腳本(緊急用):
|
||||
-- DROP INDEX IF EXISTS idx_rag_query_log_embedding;
|
||||
-- DROP INDEX IF EXISTS idx_rag_query_log_request_id;
|
||||
-- DROP INDEX IF EXISTS idx_rag_query_log_caller;
|
||||
-- DROP INDEX IF EXISTS idx_rag_query_log_queried_at;
|
||||
-- DROP TABLE IF EXISTS rag_query_log;
|
||||
-- =============================================================================
|
||||
|
||||
CREATE TABLE IF NOT EXISTS rag_query_log (
|
||||
id BIGSERIAL PRIMARY KEY,
|
||||
queried_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
|
||||
|
||||
-- 與 ai_calls.caller 同一張白名單(hermes_qa / openclaw_qa / ...)
|
||||
caller VARCHAR(64) NOT NULL,
|
||||
|
||||
-- 用戶查詢(PII 風險,限 4KB;不 normalize 以保留原始查詢樣貌)
|
||||
query_text TEXT NOT NULL,
|
||||
|
||||
-- bge-m3 embedding(1024 維,與 ai_insights.embedding 同源;可 NULL = embedding 失敗仍記錄此次嘗試)
|
||||
query_embedding VECTOR(1024),
|
||||
|
||||
-- 召回參數
|
||||
top_k INTEGER NOT NULL DEFAULT 5,
|
||||
threshold NUMERIC(4,3) NOT NULL DEFAULT 0.85,
|
||||
|
||||
-- 召回結果
|
||||
hit_count INTEGER NOT NULL DEFAULT 0,
|
||||
used_results BIGINT[], -- 命中的 ai_insights.id 陣列(軟連結,不加 FK)
|
||||
|
||||
-- 是否成功避免 LLM 呼叫(核心成本指標)
|
||||
saved_call BOOLEAN NOT NULL DEFAULT FALSE,
|
||||
|
||||
-- Telegram 👍/👎 反饋(1-5;NULL = 未反饋)
|
||||
feedback_score INTEGER,
|
||||
|
||||
-- 與 ai_calls.request_id 串鏈
|
||||
request_id VARCHAR(64),
|
||||
|
||||
-- ─────── 護欄 (對齊 critic-A11 風格) ───────
|
||||
CONSTRAINT chk_rag_threshold CHECK (
|
||||
threshold BETWEEN 0 AND 1
|
||||
),
|
||||
CONSTRAINT chk_rag_top_k CHECK (
|
||||
top_k BETWEEN 1 AND 50
|
||||
),
|
||||
CONSTRAINT chk_rag_hit_count CHECK (
|
||||
hit_count >= 0 AND hit_count <= top_k
|
||||
),
|
||||
CONSTRAINT chk_rag_query_size CHECK (
|
||||
octet_length(query_text) <= 4096
|
||||
),
|
||||
CONSTRAINT chk_rag_feedback CHECK (
|
||||
feedback_score IS NULL OR feedback_score BETWEEN 1 AND 5
|
||||
),
|
||||
-- saved_call=TRUE 必須有命中(hit_count > 0)才合理
|
||||
CONSTRAINT chk_rag_saved_consistent CHECK (
|
||||
(saved_call = FALSE) OR (hit_count > 0)
|
||||
)
|
||||
);
|
||||
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
-- 索引設計
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
-- (1) 時間範圍掃描(日報 / 命中率報表)
|
||||
CREATE INDEX IF NOT EXISTS idx_rag_query_log_queried_at
|
||||
ON rag_query_log (queried_at DESC);
|
||||
|
||||
-- (2) caller 分布(哪個入口 RAG 命中率高)
|
||||
CREATE INDEX IF NOT EXISTS idx_rag_query_log_caller
|
||||
ON rag_query_log (caller, queried_at DESC);
|
||||
|
||||
-- (3) request_id 串鏈(部分索引,sparse 不全建)
|
||||
CREATE INDEX IF NOT EXISTS idx_rag_query_log_request_id
|
||||
ON rag_query_log (request_id)
|
||||
WHERE request_id IS NOT NULL;
|
||||
|
||||
-- (4) pgvector ivfflat(cosine similarity;只索引非 NULL embedding)
|
||||
-- 注意: ivfflat 須先有資料才能正確訓練 lists;空表建索引會 fallback exact scan,
|
||||
-- Phase 11 灌入首批查詢後若效能退化,REINDEX CONCURRENTLY 重訓
|
||||
CREATE INDEX IF NOT EXISTS idx_rag_query_log_embedding
|
||||
ON rag_query_log
|
||||
USING ivfflat (query_embedding vector_cosine_ops)
|
||||
WITH (lists = 100)
|
||||
WHERE query_embedding IS NOT NULL;
|
||||
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
-- 權限
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
GRANT ALL PRIVILEGES ON rag_query_log TO momo;
|
||||
GRANT USAGE, SELECT ON SEQUENCE rag_query_log_id_seq TO momo;
|
||||
|
||||
-- 註: 90 天保留由 scheduler 任務執行(與 ai_calls 對齊):
|
||||
-- DELETE FROM rag_query_log WHERE queried_at < NOW() - INTERVAL '90 days';
|
||||
-- 建議 03:30 跑(ai_calls 03:00 之後),避免 IO 尖峰
|
||||
|
||||
DO $$
|
||||
BEGIN
|
||||
RAISE NOTICE 'Migration 027 done: rag_query_log + 4 indexes (ivfflat 1024d) (Operation Ollama-First v5.0 P11)';
|
||||
END $$;
|
||||
169
migrations/028_create_learning_episodes.sql
Normal file
169
migrations/028_create_learning_episodes.sql
Normal file
@@ -0,0 +1,169 @@
|
||||
-- =============================================================================
|
||||
-- Migration 028: learning_episodes — 蒸餾池 / 知識庫前哨
|
||||
-- Operation Ollama-First v5.0 — Phase 11
|
||||
-- 日期: 2026-05-03 台北
|
||||
-- 對應戰役: ADR-029(Hermes-First)+ Phase 11 PromotionGate 4 階段過濾
|
||||
-- =============================================================================
|
||||
-- 說明:
|
||||
-- LLM/MCP 結果先寫入 learning_episodes(蒸餾池),過 4 階段 PromotionGate
|
||||
-- 才晉升 ai_insights(知識庫主檔)。設計目的:
|
||||
-- - 隔離未驗證內容,避免直接污染 RAG 召回語料
|
||||
-- - 保留 raw + distilled,方便事後重訓
|
||||
-- - 高權重(>=0.8)走人工驗收,低權重走自動晉升
|
||||
--
|
||||
-- PromotionGate 狀態機:
|
||||
-- pending
|
||||
-- ├─[Stage 1: quality<0.7]→ rejected_quality
|
||||
-- ├─[Stage 2: 規則檢測幻覺]→ rejected_hallucination
|
||||
-- ├─[Stage 3: 與既有 insight cosine>0.95]→ rejected_duplicate
|
||||
-- ├─[Stage 4a: weight<0.8 + 過 1-3]→ approved → 寫 ai_insights → insight_id 回填
|
||||
-- └─[Stage 4b: weight>=0.8]→ awaiting_review → Telegram 推播
|
||||
-- ├─[人工 👍]→ approved
|
||||
-- ├─[人工 👎]→ rejected_human
|
||||
-- └─[24h 無反饋]→ expired (weight 降為 0.5 重走 Stage 4a)
|
||||
--
|
||||
-- 設計決策:
|
||||
-- 1. insight_id 軟連結(不加 FK)—— ai_insights archive 不應 cascade 影響蒸餾池
|
||||
-- 2. source_table + source_id 軟連結到 ai_calls / mcp_calls,方便事後重訓溯源
|
||||
-- 3. embedding 與 rag_query_log 同 1024 維,跨表 cosine 一致
|
||||
-- 4. 不設 90 天保留(蒸餾池長期保留;approved/rejected_* 進冷儲檔由後續 ADR 定)
|
||||
-- —— 短期內暴增風險:靠 partial index + monthly archive scheduler 控制
|
||||
-- 5. promotion_status 用 VARCHAR(32) + CHECK 白名單;不上 ENUM 因新增狀態方便
|
||||
-- 6. rejected_reason CHECK 強制 rejected_* 狀態必填,避免「沒原因的拒絕」
|
||||
-- 7. human_approver 存 Telegram username 的 SHA1[:8],避免 PII 落地
|
||||
--
|
||||
-- 回滾腳本(緊急用):
|
||||
-- DROP INDEX IF EXISTS idx_le_embedding;
|
||||
-- DROP INDEX IF EXISTS idx_le_insight_id;
|
||||
-- DROP INDEX IF EXISTS idx_le_episode_type;
|
||||
-- DROP INDEX IF EXISTS idx_le_status;
|
||||
-- DROP INDEX IF EXISTS idx_le_created_at;
|
||||
-- DROP TABLE IF EXISTS learning_episodes;
|
||||
-- =============================================================================
|
||||
|
||||
CREATE TABLE IF NOT EXISTS learning_episodes (
|
||||
id BIGSERIAL PRIMARY KEY,
|
||||
created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
|
||||
|
||||
-- 來源類型
|
||||
-- mcp_result = MCP server 抓回的事實(grounding / search / db query)
|
||||
-- llm_response = LLM 生成的洞察 / 摘要(hermes_analyst / openclaw 等)
|
||||
-- user_feedback = 用戶 Telegram 直接告知的事實(高 weight,需人工確認)
|
||||
-- manual_curated = 人工手動入庫(最高 weight,跳 PromotionGate)
|
||||
episode_type VARCHAR(32) NOT NULL,
|
||||
|
||||
-- 軟連結來源(不加 FK)
|
||||
source_table VARCHAR(32), -- 'ai_calls' / 'mcp_calls' / NULL
|
||||
source_id BIGINT, -- 對應 source_table 的 id
|
||||
|
||||
-- 蒸餾後的精煉文本(≤16KB;raw 不存在此表,由 source_table 透過 source_id 回查)
|
||||
distilled_text TEXT NOT NULL,
|
||||
embedding VECTOR(1024), -- 與 ai_insights / rag_query_log 同維
|
||||
|
||||
-- 蒸餾品質評分(0-1)
|
||||
-- <0.7 → Stage 1 直接 rejected_quality
|
||||
-- >=0.7 → 進 Stage 2-3
|
||||
quality_score NUMERIC(4,3) NOT NULL DEFAULT 0.0,
|
||||
|
||||
-- 權重(影響晉升路徑)
|
||||
-- <0.8 → Stage 4a 自動晉升
|
||||
-- >=0.8 → Stage 4b 人工驗收
|
||||
weight NUMERIC(4,3) NOT NULL DEFAULT 0.5,
|
||||
|
||||
-- PromotionGate 狀態(見上方狀態機)
|
||||
promotion_status VARCHAR(32) NOT NULL DEFAULT 'pending',
|
||||
|
||||
-- 晉升結果
|
||||
insight_id BIGINT, -- 晉升後對應 ai_insights.id(軟連結,無 FK)
|
||||
rejected_reason TEXT, -- promotion_status=rejected_* 時必填
|
||||
human_approver VARCHAR(64), -- Telegram username SHA1[:8]
|
||||
reviewed_at TIMESTAMPTZ,
|
||||
|
||||
-- ─────── 護欄 (對齊 critic-A11 風格) ───────
|
||||
CONSTRAINT chk_le_quality CHECK (
|
||||
quality_score BETWEEN 0 AND 1
|
||||
),
|
||||
CONSTRAINT chk_le_weight CHECK (
|
||||
weight BETWEEN 0 AND 1
|
||||
),
|
||||
CONSTRAINT chk_le_episode_type CHECK (
|
||||
episode_type IN ('mcp_result','llm_response','user_feedback','manual_curated')
|
||||
),
|
||||
CONSTRAINT chk_le_status CHECK (
|
||||
promotion_status IN (
|
||||
'pending','approved','awaiting_review',
|
||||
'rejected_quality','rejected_hallucination','rejected_duplicate','rejected_human',
|
||||
'expired'
|
||||
)
|
||||
),
|
||||
CONSTRAINT chk_le_distilled_size CHECK (
|
||||
octet_length(distilled_text) <= 16384
|
||||
),
|
||||
CONSTRAINT chk_le_rejected_reason CHECK (
|
||||
(promotion_status NOT LIKE 'rejected_%') OR (rejected_reason IS NOT NULL)
|
||||
),
|
||||
-- approved 必須有 insight_id;其他狀態不應有
|
||||
CONSTRAINT chk_le_approved_consistent CHECK (
|
||||
(promotion_status = 'approved') = (insight_id IS NOT NULL)
|
||||
),
|
||||
-- source_table + source_id 一致性(要嘛兩個都 NULL,要嘛兩個都有)
|
||||
CONSTRAINT chk_le_source_consistent CHECK (
|
||||
(source_table IS NULL AND source_id IS NULL)
|
||||
OR (source_table IS NOT NULL AND source_id IS NOT NULL)
|
||||
),
|
||||
CONSTRAINT chk_le_source_table CHECK (
|
||||
source_table IS NULL OR source_table IN ('ai_calls','mcp_calls')
|
||||
),
|
||||
-- 人工驗收時 reviewed_at 必填
|
||||
CONSTRAINT chk_le_review_consistent CHECK (
|
||||
(human_approver IS NULL) OR (reviewed_at IS NOT NULL)
|
||||
)
|
||||
);
|
||||
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
-- 索引設計
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
-- (1) 時間範圍掃描(蒸餾池規模監控)
|
||||
CREATE INDEX IF NOT EXISTS idx_le_created_at
|
||||
ON learning_episodes (created_at DESC);
|
||||
|
||||
-- (2) 待處理佇列查詢(PromotionGate worker / 人工驗收 dashboard)
|
||||
-- partial index 縮體積:只關心 pending / awaiting_review 兩種「活躍」狀態
|
||||
CREATE INDEX IF NOT EXISTS idx_le_status
|
||||
ON learning_episodes (promotion_status, created_at DESC)
|
||||
WHERE promotion_status IN ('pending','awaiting_review');
|
||||
|
||||
-- (3) 來源類型分布報表
|
||||
CREATE INDEX IF NOT EXISTS idx_le_episode_type
|
||||
ON learning_episodes (episode_type, created_at DESC);
|
||||
|
||||
-- (4) insight_id 反查(從 ai_insights 反推蒸餾來源)
|
||||
CREATE INDEX IF NOT EXISTS idx_le_insight_id
|
||||
ON learning_episodes (insight_id)
|
||||
WHERE insight_id IS NOT NULL;
|
||||
|
||||
-- (5) pgvector ivfflat(Stage 3 重複檢測 cosine query 主用)
|
||||
CREATE INDEX IF NOT EXISTS idx_le_embedding
|
||||
ON learning_episodes
|
||||
USING ivfflat (embedding vector_cosine_ops)
|
||||
WITH (lists = 100)
|
||||
WHERE embedding IS NOT NULL;
|
||||
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
-- 權限
|
||||
-- ─────────────────────────────────────────────────────────────────────────────
|
||||
GRANT ALL PRIVILEGES ON learning_episodes TO momo;
|
||||
GRANT USAGE, SELECT ON SEQUENCE learning_episodes_id_seq TO momo;
|
||||
|
||||
-- 註: expired 狀態降權 worker(24h 無反饋)由 scheduler 跑:
|
||||
-- UPDATE learning_episodes
|
||||
-- SET promotion_status='expired', weight=0.5
|
||||
-- WHERE promotion_status='awaiting_review'
|
||||
-- AND created_at < NOW() - INTERVAL '24 hours';
|
||||
-- 之後由 PromotionGate Stage 4a 重跑該批 expired 走自動晉升路徑。
|
||||
|
||||
DO $$
|
||||
BEGIN
|
||||
RAISE NOTICE 'Migration 028 done: learning_episodes + 5 indexes + 9 CHECK constraints (Operation Ollama-First v5.0 P11)';
|
||||
END $$;
|
||||
Reference in New Issue
Block a user