ewoooc/services/openclaw_learning_service.py

import json
import math
import threading
import time
from datetime import datetime, timezone
from sqlalchemy import text
from services.logger_manager import SystemLogger
from database.manager import get_session
from database.ai_models import AIInsight
from services.ollama_service import ollama_service

sys_log = SystemLogger("OCLearn").get_logger()

# =====================================================================
# ADR-007 Step 4 (2026-04-19): DB 持久化 Embedding Retry Queue
#
# 改造原因：原記憶體 Queue 重啟會遺失未處理項目，違反 ADR-007「雙寫必達」。
# 新架構：入列寫 embedding_retry_queue 表 → Hermes worker 每分鐘輪詢批次處理。
# =====================================================================

EMBED_POLL_INTERVAL_SEC = 60         # worker 輪詢間隔
EMBED_BATCH_SIZE = 10                # 單次最多處理筆數
EMBED_MAX_ATTEMPTS = 5               # 超過則標記 failed
DECAY_RATE = 0.01                    # ADR-005：半衰期約 70 天


def _enqueue_embedding(target_table: str, target_id: int, text_content: str,
                       model: str = "bge-m3:latest") -> bool:
    """將待 embed 項目寫入 DB retry queue（持久化）"""
    session = get_session()
    try:
        session.execute(
            text("""
                INSERT INTO embedding_retry_queue
                    (target_table, target_id, text_content, model, status, created_at)
                VALUES (:t, :i, :txt, :m, 'pending', :now)
            """),
            {
                "t": target_table,
                "i": target_id,
                "txt": text_content,
                "m": model,
                "now": datetime.now(),
            },
        )
        session.commit()
        return True
    except Exception as e:
        session.rollback()
        sys_log.warning(f"[OCLearn] enqueue embedding 失敗 (可能表尚未建立): {e}")
        return False
    finally:
        session.close()


def _process_one_embedding(row_id: int, target_table: str, target_id: int,
                           text_content: str, model: str) -> bool:
    """處理單筆 embedding，成功寫回目標表，失敗累加 attempts"""
    session = get_session()
    try:
        session.execute(
            text("UPDATE embedding_retry_queue SET status='processing', updated_at=:now WHERE id=:id"),
            {"now": datetime.now(), "id": row_id},
        )
        session.commit()

        vec = ollama_service.generate_embedding(text_content, model=model)
        if not vec:
            raise RuntimeError("embedding 回傳空值")

        vec_str = str(vec)
        session.execute(
            text(f"UPDATE {target_table} SET embedding = :vec WHERE id = :id"),
            {"vec": vec_str, "id": target_id},
        )
        session.execute(
            text("""
                UPDATE embedding_retry_queue
                   SET status='done', processed_at=:now, updated_at=:now
                 WHERE id=:id
            """),
            {"now": datetime.now(), "id": row_id},
        )
        session.commit()
        sys_log.debug(f"[OCLearn] embedding 寫入成功 {target_table}#{target_id}")
        return True

    except Exception as e:
        session.rollback()
        try:
            session.execute(
                text("""
                    UPDATE embedding_retry_queue
                       SET attempts = attempts + 1,
                           last_error = :err,
                           status = CASE WHEN attempts + 1 >= :max THEN 'failed' ELSE 'pending' END,
                           updated_at = :now
                     WHERE id = :id
                """),
                {
                    "err": str(e)[:500],
                    "max": EMBED_MAX_ATTEMPTS,
                    "now": datetime.now(),
                    "id": row_id,
                },
            )
            session.commit()
        except Exception as e2:
            session.rollback()
            sys_log.error(f"[OCLearn] 更新 retry attempts 失敗: {e2}")
        sys_log.warning(f"[OCLearn] embedding 失敗 (attempt 累加): {e}")
        return False
    finally:
        session.close()


def _embedding_worker_loop():
    """背景執行緒：輪詢 embedding_retry_queue，批次處理 pending 項目"""
    sys_log.info("[OCLearn] Hermes Embedding Worker (DB-backed) 啟動")
    while True:
        try:
            session = get_session()
            try:
                rows = session.execute(
                    text("""
                        SELECT id, target_table, target_id, text_content, model
                          FROM embedding_retry_queue
                         WHERE status = 'pending'
                           AND attempts < :max
                         ORDER BY created_at
                         LIMIT :lim
                    """),
                    {"max": EMBED_MAX_ATTEMPTS, "lim": EMBED_BATCH_SIZE},
                ).fetchall()
            finally:
                session.close()

            for row in rows:
                _process_one_embedding(
                    row[0], row[1], row[2], row[3], row[4] or "bge-m3:latest"
                )
        except Exception as e:
            sys_log.warning(f"[OCLearn] worker 輪詢略過 (表可能未建立): {e}")

        time.sleep(EMBED_POLL_INTERVAL_SEC)


# 啟動背景 Worker（daemon，Python 進程結束會自然回收）
threading.Thread(target=_embedding_worker_loop, daemon=True).start()


# =====================================================================
# ADR-007 核心：store_insight 雙寫（DB + 排程 embedding）
# =====================================================================

def store_insight(insight_type: str, content: str, period: str = None,
                  product_sku: str = None, metadata: dict = None,
                  ai_model: 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
    - 向量化：寫入 embedding_retry_queue 供 Hermes worker 處理（持久化）
    """
    session = get_session()
    try:
        meta_str = json.dumps(metadata, ensure_ascii=False) if metadata else None

        existing = None
        if period and insight_type:
            q = session.query(AIInsight).filter_by(insight_type=insight_type, period=period)
            if product_sku:
                q = q.filter_by(product_sku=product_sku)
            existing = q.first()

        if existing:
            existing.content = content
            if meta_str:
                existing.metadata_json = meta_str
            existing.updated_at = datetime.now()
            session.commit()
            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(),
            )
            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）
        if ai_model:
            try:
                session.execute(
                    text("UPDATE ai_insights SET ai_model = :m WHERE id = :i"),
                    {"m": ai_model, "i": insight_id},
                )
                session.commit()
            except Exception:
                session.rollback()

        # 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)

        return insight_id

    except Exception as e:
        session.rollback()
        sys_log.error(f"[OCLearn] store_insight 錯誤: {e}")
        return None
    finally:
        session.close()


# =====================================================================
# ADR-005: 時間衰減品質分數
# =====================================================================

def compute_effective_score(base_score: float, created_at: datetime,
                            decay_rate: float = DECAY_RATE) -> float:
    """Effective_Score = base × exp(-λ × days_since_created)"""
    if not created_at:
        return base_score
    if created_at.tzinfo is None:
        created_at = created_at.replace(tzinfo=timezone.utc)
    days = (datetime.now(timezone.utc) - created_at).days
    return base_score * math.exp(-decay_rate * max(days, 0))


# =====================================================================
# RAG 檢索（套時間衰減）
# =====================================================================

def build_rag_context(query: str, insight_type: str = None, period: str = None,
                      top_k: int = 5) -> str:
    """
    RAG 上下文檢索（Cache-aside 精確命中 + 未來語意檢索）
    2026-04-19: 加入時間衰減重排（ADR-005）
    """
    session = get_session()
    try:
        q = session.query(AIInsight)
        if insight_type:
            q = q.filter_by(insight_type=insight_type)
        if period:
            q = q.filter_by(period=period)

        if not period:
            q = q.order_by(AIInsight.created_at.desc()).limit(20)

        results = q.all()
        if not results:
            return ""

        # 套用時間衰減重排（無 avg_quality 欄位時，預設 base=0.5）
        ranked = []
        for r in results:
            base = getattr(r, "avg_quality", None) or 0.5
            exempt = getattr(r, "decay_exempt", False)
            effective = base if exempt else compute_effective_score(base, r.created_at)
            ranked.append((effective, r))
        ranked.sort(key=lambda x: x[0], reverse=True)

        top = ranked[:top_k]
        context_parts = []
        for score, r in top:
            p_tag = f"[{r.period}]" if r.period else "[歷史紀錄]"
            context_parts.append(
                f"{p_tag} {r.insight_type} (分數={score:.2f}):\n{r.content}"
            )

        sys_log.info(f"[OCLearn] RAG context: {len(top)}/{len(results)} 筆（時間衰減重排）")
        return "\n\n---\n\n".join(context_parts)
    except Exception as e:
        sys_log.error(f"[OCLearn] build_rag_context 錯誤: {e}")
        return ""
    finally:
        session.close()


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()
    try:
        start_dt = datetime.strptime(f"{start_date} 00:00:00", "%Y-%m-%d %H:%M:%S")
        end_dt = datetime.strptime(f"{end_date} 23:59:59", "%Y-%m-%d %H:%M:%S")

        results = (
            session.query(AIInsight)
            .filter(AIInsight.created_at >= start_dt)
            .filter(AIInsight.created_at <= end_dt)
            .order_by(AIInsight.created_at.asc())
            .all()
        )
        if not results:
            return ""

        parts = []
        for r in results:
            dt_str = r.created_at.strftime("%Y-%m-%d %H:%M") if r.created_at else ""
            p_tag = f"[{r.period if r.period else dt_str}]"
            parts.append(f"{p_tag} 洞察類型({r.insight_type}):\n{r.content}")

        sys_log.info(
            f"[OCLearn] 週報 RAG: {len(results)} 筆（{start_date}~{end_date}）"
        )
        return "\n\n---\n\n".join(parts)
    except Exception as e:
        sys_log.error(f"[OCLearn] build_rag_context_by_date 錯誤: {e}")
        return ""
    finally:
        session.close()


# =====================================================================
# 其他 API
# =====================================================================

def store_conversation(user_id: int, user_message: str, bot_response: str) -> bool:
    """用戶對話沉澱（走 store_insight 雙寫）"""
    try:
        store_insight(
            insight_type="chat_history",
            content=f"User: {user_message}\nBot: {bot_response}",
            metadata={"user_id": user_id},
        )
        return True
    except Exception as e:
        sys_log.error(f"[OCLearn] store_conversation 錯誤: {e}")
        return False


def update_feedback(insight_id: int, rating: int, comment: str = None) -> bool:
    """用戶回饋：累加 feedback_up/down 並重算 avg_quality（Migration 010 後生效）"""
    session = get_session()
    try:
        col = "feedback_up" if rating >= 4 else "feedback_down"
        session.execute(
            text(f"UPDATE ai_insights SET {col} = COALESCE({col},0) + 1 WHERE id = :id"),
            {"id": insight_id},
        )
        # 重算 avg_quality = up / (up+down)
        session.execute(
            text("""
                UPDATE ai_insights
                   SET avg_quality = CASE
                        WHEN COALESCE(feedback_up,0) + COALESCE(feedback_down,0) = 0 THEN 0.5
                        ELSE COALESCE(feedback_up,0)::float
                             / (COALESCE(feedback_up,0) + COALESCE(feedback_down,0))
                       END
                 WHERE id = :id
            """),
            {"id": insight_id},
        )
        session.commit()
        sys_log.info(f"[OCLearn] Feedback {insight_id}: rating={rating}")
        return True
    except Exception as e:
        session.rollback()
        sys_log.warning(f"[OCLearn] update_feedback 略過 (Migration 010 未執行?): {e}")
        return False
    finally:
        session.close()


def get_learning_stats() -> dict:
    """學習系統總攬指標"""
    session = get_session()
    try:
        count = session.query(AIInsight).count()
        try:
            pending = session.execute(
                text("SELECT COUNT(*) FROM embedding_retry_queue WHERE status='pending'")
            ).scalar()
        except Exception:
            pending = None
        return {
            "total_insights": count,
            "embedding_queue_pending": pending,
            "status": "active",
        }
    except Exception as e:
        return {"total_insights": 0, "status": "error", "error": str(e)}
    finally:
        session.close()


# =====================================================================
# ADR-005 Step-2: 每日批次維護（03:00 去重 / 04:00 品質分數重算）
# =====================================================================

def run_dedup_batch() -> dict:
    """
    03:00 批次：去除同日同 SKU 同 insight_type 的重複洞察。
    策略：保留 avg_quality 最高者（tie-break: 最新 id），其餘設 status='archived'。
    回傳 {"archived": int, "scanned": int}
    """
    session = get_session()
    archived = 0
    scanned = 0
    try:
        dupes = session.execute(text("""
            SELECT insight_type, product_sku, period, COUNT(*) AS cnt
              FROM ai_insights
             WHERE status NOT IN ('archived', 'relearn')
               AND product_sku IS NOT NULL
             GROUP BY insight_type, product_sku, period
            HAVING COUNT(*) > 1
        """)).fetchall()

        for row in dupes:
            scanned += row[3]
            keep = session.execute(text("""
                SELECT id FROM ai_insights
                 WHERE insight_type = :t AND product_sku = :sku
                   AND period IS NOT DISTINCT FROM :p
                   AND status NOT IN ('archived', 'relearn')
                 ORDER BY avg_quality DESC NULLS LAST, id DESC
                 LIMIT 1
            """), {"t": row[0], "sku": row[1], "p": row[2]}).scalar()

            if keep is None:
                continue

            result = session.execute(text("""
                UPDATE ai_insights
                   SET status = 'archived', updated_at = CURRENT_TIMESTAMP
                 WHERE insight_type = :t AND product_sku = :sku
                   AND period IS NOT DISTINCT FROM :p
                   AND status NOT IN ('archived', 'relearn')
                   AND id != :keep_id
            """), {"t": row[0], "sku": row[1], "p": row[2], "keep_id": keep})
            archived += result.rowcount

        session.commit()
        sys_log.info(f"[OCLearn] 去重批次完成：掃描 {scanned} 筆，歸檔 {archived} 筆")
        return {"archived": archived, "scanned": scanned}
    except Exception as e:
        session.rollback()
        sys_log.error(f"[OCLearn] 去重批次失敗: {e}")
        return {"archived": 0, "scanned": 0, "error": str(e)}
    finally:
        session.close()


def run_quality_rescore_batch() -> dict:
    """
    04:00 批次：對所有 approved/relearn 狀態的 ai_insights 套用時間衰減公式。
    decay_exempt=True 的記錄跳過衰減（永久知識）。
    relearn 狀態額外懲罰 20%。分數 < 0.05 自動歸檔。
    回傳 {"updated": int, "relearn_reset": int}
    """
    session = get_session()
    updated = 0
    relearn_reset = 0
    try:
        rows = session.execute(text("""
            SELECT id, avg_quality, created_at, decay_exempt, status
              FROM ai_insights
             WHERE status IN ('approved', 'relearn')
        """)).fetchall()

        for row in rows:
            row_id, base_q, created_at, exempt, status = row
            base_q = base_q or 0.5

            if exempt:
                continue

            if created_at and created_at.tzinfo is None:
                created_at = created_at.replace(tzinfo=timezone.utc)

            new_score = compute_effective_score(base_q, created_at)
            if status == "relearn":
                new_score *= 0.8

            session.execute(text("""
                UPDATE ai_insights
                   SET avg_quality = :q, updated_at = CURRENT_TIMESTAMP
                 WHERE id = :id
            """), {"q": round(new_score, 4), "id": row_id})
            updated += 1

            if new_score < 0.05:
                session.execute(text(
                    "UPDATE ai_insights SET status = 'archived' WHERE id = :id"
                ), {"id": row_id})
                if status == "relearn":
                    relearn_reset += 1

        session.commit()
        sys_log.info(
            f"[OCLearn] 品質分數重算完成：更新 {updated} 筆，"
            f"低分自動歸檔含 relearn {relearn_reset} 筆"
        )
        return {"updated": updated, "relearn_reset": relearn_reset}
    except Exception as e:
        session.rollback()
        sys_log.error(f"[OCLearn] 品質分數重算批次失敗: {e}")
        return {"updated": 0, "relearn_reset": 0, "error": str(e)}
    finally:
        session.close()