From 00c80c3beca77769ad7f2904b00592844538acb5 Mon Sep 17 00:00:00 2001
From: jimi <1847930177@qq.com>
Date: Sun, 1 Mar 2026 17:09:05 +0800
Subject: [PATCH] feat: ai-first intent detection with keyword fallback

---
 core/collection_intent_helpers.py |  5 ++-
 core/context_helpers.py           |  9 +++---
 core/websocket_client.py          | 54 ++++++++++++++++++++++++-------
 tests/test_intent_analyzer.py     | 24 ++++++++++++++
 utils/intent_analyzer.py          | 49 +++++++++++++++++++++++++---
 5 files changed, 117 insertions(+), 24 deletions(-)
 create mode 100644 tests/test_intent_analyzer.py

diff --git a/core/collection_intent_helpers.py b/core/collection_intent_helpers.py
index a450703..0bd1888 100644
--- a/core/collection_intent_helpers.py
+++ b/core/collection_intent_helpers.py
@@ -148,9 +148,8 @@ def is_batch_finish_intent(text: str, state: Any, has_incoming_urls: bool) -> bo
         return False
 
     try:
-        from utils.intent_analyzer import detect_intent_embedding, detect_intent_keywords
-
-        intent = detect_intent_embedding(text) or detect_intent_keywords(text)
+        from utils.intent_analyzer import detect_intent
+        intent = detect_intent(text).intent
     except Exception:
         intent = ""
     if intent in ("询价", "砍价"):
diff --git a/core/context_helpers.py b/core/context_helpers.py
index 3a2d53b..c42767d 100644
--- a/core/context_helpers.py
+++ b/core/context_helpers.py
@@ -210,15 +210,16 @@ def get_conversation_context(customer_id: str, acc_id: str = "", limit: int = 12
 def get_intent_emotion_hint(msg: str) -> str:
     """语义匹配：意图/情绪识别，注入提示。EMBEDDING_MODEL 未配置时用关键词。"""
     try:
-        from utils.intent_analyzer import detect_emotion_embedding, detect_intent_embedding, detect_intent_keywords
+        from utils.intent_analyzer import detect_emotion_embedding, detect_intent
 
-        intent = detect_intent_embedding(msg)
-        if not intent:
-            intent = detect_intent_keywords(msg)
+        decision = detect_intent(msg)
+        intent = decision.intent
         emotion = detect_emotion_embedding(msg) if os.getenv("EMBEDDING_MODEL") else None
         parts = []
         if intent:
             parts.append(f"意图:{intent}")
+            if decision.source:
+                parts.append(f"意图来源:{decision.source}")
         if emotion:
             parts.append(f"情绪:{emotion}")
         if parts:
diff --git a/core/websocket_client.py b/core/websocket_client.py
index a822228..8dd7a85 100755
--- a/core/websocket_client.py
+++ b/core/websocket_client.py
@@ -396,7 +396,14 @@ class QingjianAPIClient:
                     elif self._is_shop_card(data):
                         # 进店卡片：有历史对话就不回复，没有才打招呼（Gemini 已在上面统一跳过）
                         cid = data.get('from_id', '')
-                        if self._has_chat_history(cid):
+                        acc_id = data.get('acc_id', '')
+                        residual_text = self._extract_customer_text_from_shop_card_msg(data.get('msg', ''))
+                        if residual_text:
+                            logger.info(f"[{self.get_time()}] 进店卡片携带客户文本，转普通消息处理: {residual_text}")
+                            patched = dict(data)
+                            patched['msg'] = residual_text
+                            await self._debounce_agent_reply(patched)
+                        elif self._has_chat_history(cid, acc_id=acc_id):
                             logger.info(f"[{self.get_time()}] 进店卡片（已有记录），跳过")
                         else:
                             logger.info(f"[{self.get_time()}] 进店卡片（新客户），发送问候")
@@ -505,14 +512,23 @@ class QingjianAPIClient:
         if self._msg_has_image_url(text):
             return "image"
         try:
-            from utils.intent_analyzer import detect_intent_keywords
-            intent = detect_intent_keywords(text)
+            from utils.intent_analyzer import detect_intent
+            decision = detect_intent(text)
+            intent = decision.intent
+            if intent:
+                self._activity_log(
+                    "debounce_intent_detected",
+                    intent=intent,
+                    source=decision.source,
+                    score=round(float(decision.score or 0.0), 4),
+                    msg=text[:120],
+                )
         except Exception:
             intent = ""
         if intent:
             return intent
         lower = text.lower()
-        if any(k in lower for k in ["报价", "多少钱", "价格", "贵", "优惠"]):
+        if any(k in lower for k in ["报价", "多少钱", "价格", "贵", "优惠", "收费", "怎么收费", "咋收费"]):
             return "询价"
         if any(k in lower for k in ["做一下", "改一下", "需求", "门头", "上面的字", "处理"]):
             return "修改"
@@ -1222,17 +1238,31 @@ class QingjianAPIClient:
         msg = self.to_chinese(data.get('msg', ''))
         return msg.startswith('[进店卡片]') or '我想咨询你们店的这个商品' in msg
 
-    def _has_chat_history(self, customer_id: str) -> bool:
-        """判断该客户是否已有聊天记录（内存历史或数据库均可）"""
+    def _extract_customer_text_from_shop_card_msg(self, msg: str) -> str:
+        """从“进店卡片+文本”混合消息里提取客户真实文本。"""
+        text = self.to_chinese(msg or "").strip()
+        if not text:
+            return ""
+        parts = [p.strip() for p in text.split("#*#") if p and p.strip()]
+        kept = []
+        for part in parts:
+            if part.startswith("[进店卡片]") or "我想咨询你们店的这个商品" in part:
+                continue
+            kept.append(part)
+        if kept:
+            return " ".join(kept).strip()
+        stripped = re.sub(r"\[进店卡片\][^\n\r]*", "", text).strip()
+        stripped = stripped.replace("我想咨询你们店的这个商品", "").strip("，。,#* ")
+        return stripped
+
+    def _has_chat_history(self, customer_id: str, acc_id: str = "") -> bool:
+        """判断该客户在当前店铺是否已有聊天记录。"""
         if not customer_id:
             return False
-        # 先查内存对话历史（最快）
-        if customer_id in self.agent.message_histories and self.agent.message_histories[customer_id]:
-            return True
-        # 再查数据库（重启后仍有记录）
+        # 按店铺+客户查数据库，避免跨店串历史导致错误跳过。
         try:
-            from db.chat_log_db import get_conversation
-            msgs = get_conversation(customer_id, limit=1)
+            from db.chat_log_db import get_recent_conversation
+            msgs = get_recent_conversation(customer_id, acc_id=acc_id, limit=1)
             return len(msgs) > 0
         except Exception:
             return False
diff --git a/tests/test_intent_analyzer.py b/tests/test_intent_analyzer.py
new file mode 100644
index 0000000..582fe70
--- /dev/null
+++ b/tests/test_intent_analyzer.py
@@ -0,0 +1,24 @@
+import unittest
+
+from utils.intent_analyzer import detect_intent
+
+
+class IntentAnalyzerTests(unittest.TestCase):
+    def test_keyword_fallback_for_price(self):
+        d = detect_intent("这个怎么收费")
+        self.assertEqual(d.intent, "询价")
+        self.assertEqual(d.source, "keyword")
+
+    def test_keyword_fallback_for_greeting(self):
+        d = detect_intent("你好 在吗")
+        self.assertEqual(d.intent, "打招呼")
+        self.assertEqual(d.source, "keyword")
+
+    def test_unknown_intent(self):
+        d = detect_intent("abc123")
+        self.assertEqual(d.intent, "")
+        self.assertIn(d.source, ("none", ""))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/utils/intent_analyzer.py b/utils/intent_analyzer.py
index a2f2513..70fc499 100755
--- a/utils/intent_analyzer.py
+++ b/utils/intent_analyzer.py
@@ -5,7 +5,8 @@
 """
 import os
 import logging
-from typing import Optional, Tuple
+from dataclasses import dataclass
+from typing import Optional
 
 logger = logging.getLogger(__name__)
 
@@ -31,6 +32,13 @@ EMOTION_TEMPLATES = {
 
 _template_embeddings: dict = {}
 
+
+@dataclass
+class IntentDecision:
+    intent: str = ""
+    source: str = "none"  # embedding / keyword / none
+    score: float = 0.0
+
 def _get_embedding(text: str, cache_key: str = None) -> Optional[list]:
     """调用 embedding API，失败返回 None。cache_key 用于缓存模板向量"""
     model = os.getenv("EMBEDDING_MODEL", "")
@@ -66,10 +74,16 @@ def _cosine_sim(a: list, b: list) -> float:
 
 
 def detect_intent_embedding(msg: str) -> Optional[str]:
-    """用 embedding 检测意图，未配置或失败返回 None"""
+    """用 embedding 检测意图，未配置或失败返回 None。"""
+    decision = detect_intent_embedding_decision(msg)
+    return decision.intent or None
+
+
+def detect_intent_embedding_decision(msg: str) -> IntentDecision:
+    """返回 embedding 意图决策（含分值）。"""
     msg_emb = _get_embedding(msg)
     if not msg_emb:
-        return None
+        return IntentDecision()
     best_intent, best_score = "", 0.0
     for intent, template in INTENT_TEMPLATES.items():
         tpl_emb = _get_embedding(template, cache_key=f"intent_{intent}")
@@ -79,7 +93,9 @@ def detect_intent_embedding(msg: str) -> Optional[str]:
         if sim > best_score:
             best_score = sim
             best_intent = intent
-    return best_intent if best_score > 0.6 else None
+    if best_score > 0.6:
+        return IntentDecision(intent=best_intent, source="embedding", score=float(best_score))
+    return IntentDecision()
 
 
 def detect_emotion_embedding(msg: str) -> Optional[str]:
@@ -112,9 +128,32 @@ def detect_intent_keywords(msg: str) -> str:
         return "砍价"
     if any(k in m for k in ["改", "修改", "不满意"]):
         return "修改"
-    if any(k in m for k in ["多少钱", "价格", "报价", "多钱"]):
+    if any(k in m for k in ["多少钱", "价格", "报价", "多钱", "收费", "怎么收费", "咋收费"]):
         return "询价"
     if any(k in m for k in ["在吗", "你好", "有人"]):
         return "打招呼"
     return ""
 
+
+def detect_intent(msg: str) -> IntentDecision:
+    """
+    AI 意图判定 + 规则兜底：
+    1) 有 embedding 配置时先走 embedding。
+    2) 失败/低置信时回退关键词规则。
+    """
+    text = (msg or "").strip()
+    if not text:
+        return IntentDecision()
+
+    try:
+        emb_decision = detect_intent_embedding_decision(text)
+    except Exception:
+        emb_decision = IntentDecision()
+    if emb_decision.intent:
+        return emb_decision
+
+    kw_intent = detect_intent_keywords(text)
+    if kw_intent:
+        return IntentDecision(intent=kw_intent, source="keyword", score=0.0)
+    return IntentDecision()
+