補齊自癒稽核與 OpenClaw 記憶閉環

services/openclaw_learning_service.py

@@ -121,6 +121,21 @@ def _embedding_worker_loop():
         try:
             session = get_session()
             try:
+                session.execute(
+                    text("""
+                        UPDATE embedding_retry_queue
+                           SET status = 'pending',
+                               updated_at = :now,
+                               last_error = COALESCE(last_error, '') || ' | reset stale processing'
+                         WHERE status = 'processing'
+                           AND updated_at < :cutoff
+                    """),
+                    {
+                        "now": datetime.now(),
+                        "cutoff": datetime.fromtimestamp(time.time() - 15 * 60),
+                    },
+                )
+                session.commit()
                 rows = session.execute(
                     text("""
                         SELECT id, target_table, target_id, text_content, model
@@ -155,7 +170,8 @@ threading.Thread(target=_embedding_worker_loop, daemon=True).start()
 
 def store_insight(insight_type: str, content: str, period: str = None,
                   product_sku: str = None, metadata: dict = None,
-                  ai_model: str = None) -> int:
+                  ai_model: str = None, confidence: float = None,
+                  created_by: str = None, status: str = None) -> int:
     """
     將 AI 產出存入 ai_insights 表並排程向量化。
     - Cache-aside：同 insight_type + period + product_sku 已存在則覆蓋
@@ -181,26 +197,44 @@ def store_insight(insight_type: str, content: str, period: str = None,
             insight_id = existing.id
             sys_log.info(f"[OCLearn] 更新 insight_type={insight_type} period={period}")
         else:
-            new_insight = AIInsight(
-                insight_type=insight_type,
-                period=period,
-                product_sku=product_sku,
-                content=content,
-                metadata_json=meta_str,
-                created_at=datetime.now(),
-                updated_at=datetime.now(),
-            )
+            insight_kwargs = {
+                "insight_type": insight_type,
+                "period": period,
+                "product_sku": product_sku,
+                "content": content,
+                "metadata_json": meta_str,
+                "status": status or "approved",
+                "created_at": datetime.now(),
+                "updated_at": datetime.now(),
+            }
+            if confidence is not None:
+                insight_kwargs["confidence"] = confidence
+            if created_by:
+                insight_kwargs["created_by"] = created_by
+            new_insight = AIInsight(**insight_kwargs)
             session.add(new_insight)
             session.commit()
             insight_id = new_insight.id
             sys_log.info(f"[OCLearn] 新增 insight_type={insight_type} period={period}")
 
-        # 若已有 ai_model 欄位則寫入（Migration 010 後可用，先試再容錯）
+        # 若已有進階欄位則寫入（Migration 010/015 後可用，先試再容錯）
+        update_fields = {}
         if ai_model:
+            update_fields["ai_model"] = ai_model
+        if confidence is not None:
+            update_fields["confidence"] = confidence
+        if created_by:
+            update_fields["created_by"] = created_by
+        if status:
+            update_fields["status"] = status
+        if update_fields:
             try:
+                assignments = ", ".join(f"{key} = :{key}" for key in update_fields)
+                params = dict(update_fields)
+                params["i"] = insight_id
                 session.execute(
-                    text("UPDATE ai_insights SET ai_model = :m WHERE id = :i"),
-                    {"m": ai_model, "i": insight_id},
+                    text(f"UPDATE ai_insights SET {assignments}, updated_at = :now WHERE id = :i"),
+                    {**params, "now": datetime.now()},
                 )
                 session.commit()
             except Exception:
@@ -208,7 +242,28 @@ def store_insight(insight_type: str, content: str, period: str = None,
 
         # 推入 DB retry queue（持久化）
         embed_target_text = f"{insight_type} ({period or ''}): {content}"
-        _enqueue_embedding("ai_insights", insight_id, embed_target_text)
+        embedding_queued = _enqueue_embedding("ai_insights", insight_id, embed_target_text)
+        if not embedding_queued:
+            try:
+                metadata_payload = metadata.copy() if isinstance(metadata, dict) else {}
+                metadata_payload["embedding_queue_error"] = True
+                session.execute(
+                    text("""
+                        UPDATE ai_insights
+                           SET metadata_json = :meta,
+                               updated_at = :now
+                         WHERE id = :id
+                    """),
+                    {
+                        "meta": json.dumps(metadata_payload, ensure_ascii=False),
+                        "now": datetime.now(),
+                        "id": insight_id,
+                    },
+                )
+                session.commit()
+            except Exception as meta_err:
+                session.rollback()
+                sys_log.warning(f"[OCLearn] embedding queue error metadata 寫入失敗: {meta_err}")
 
         return insight_id
 
@@ -247,6 +302,10 @@ def build_rag_context(query: str, insight_type: str = None, period: str = None,
     """
     session = get_session()
     try:
+        semantic_context = _build_semantic_rag_context(session, query, insight_type, period, top_k)
+        if semantic_context:
+            return semantic_context
+
         q = session.query(AIInsight)
         if insight_type:
             q = q.filter_by(insight_type=insight_type)
@@ -288,6 +347,50 @@ def build_rag_context(query: str, insight_type: str = None, period: str = None,
         session.close()
 
 
+def _build_semantic_rag_context(session, query: str, insight_type: str = None,
+                                period: str = None, top_k: int = 5) -> str:
+    if not query:
+        return ""
+    try:
+        vec = ollama_service.generate_embedding(query)
+        if not vec:
+            return ""
+        filters = ["embedding IS NOT NULL", "status IN ('approved', 'active', 'executed')"]
+        params = {"qvec": str(vec), "lim": top_k}
+        if insight_type:
+            filters.append("insight_type = :insight_type")
+            params["insight_type"] = insight_type
+        if period:
+            filters.append("period = :period")
+            params["period"] = period
+        rows = session.execute(
+            text(f"""
+                SELECT id, insight_type, period, content, avg_quality, decay_exempt, created_at,
+                       embedding <=> CAST(:qvec AS vector) AS distance
+                  FROM ai_insights
+                 WHERE {' AND '.join(filters)}
+                 ORDER BY distance ASC, created_at DESC
+                 LIMIT :lim
+            """),
+            params,
+        ).fetchall()
+        if not rows:
+            return ""
+        parts = []
+        for row in rows:
+            base = row.avg_quality if row.avg_quality is not None else 0.5
+            effective = base if row.decay_exempt else compute_effective_score(base, row.created_at)
+            p_tag = f"[{row.period}]" if row.period else "[語意記憶]"
+            parts.append(
+                f"{p_tag} {row.insight_type} (語意距離={row.distance:.3f}, 分數={effective:.2f}):\n{row.content}"
+            )
+        sys_log.info(f"[OCLearn] semantic RAG context: {len(parts)} 筆")
+        return "\n\n---\n\n".join(parts)
+    except Exception as exc:
+        sys_log.debug(f"[OCLearn] semantic RAG fallback to decay ranking: {exc}")
+        return ""
+
+
 def build_rag_context_by_date(start_date: str, end_date: str) -> str:
     """週報 RAG：依日期區間過濾 ai_insights"""
     session = get_session()