fix: 优化后端数据更新机制

crawler/extractor_ai.py

@@ -15,31 +15,40 @@ CLEANER_AI_DISABLED = os.environ.get("CLEANER_AI_DISABLED", "0") == "1"
 OLLAMA_MODEL = os.environ.get("OLLAMA_MODEL", "llama3.1")
 
 
-def _call_ollama_extract(text: str, timeout: int = 10) -> Optional[Dict[str, Any]]:
-    """调用 Ollama 提取结构化数据。输出 JSON，仅包含新闻中可明确推断的字段"""
+# 用于 AI 提取的原文最大长度（有正文时取更长以提取精确数据）
+EXTRACT_TEXT_MAX_LEN = int(os.environ.get("EXTRACT_TEXT_MAX_LEN", "4000"))
+
+
+def _call_ollama_extract(text: str, timeout: int = 15) -> Optional[Dict[str, Any]]:
+    """调用 Ollama 从新闻全文/摘要中提取精确结构化数据，仅填写报道中明确给出的数字与事实。"""
     if CLEANER_AI_DISABLED or not text or len(str(text).strip()) < 10:
         return None
     try:
         import requests
-        prompt = f"""从以下美伊/中东新闻中提取可推断的数值，输出 JSON，仅包含有明确依据的字段。无依据则省略该字段。
+        raw = str(text).strip()[:EXTRACT_TEXT_MAX_LEN]
+        prompt = f"""从以下美伊/中东新闻**全文或摘要**中，提取**报道明确给出的数字与事实**，输出 JSON。规则：
+1. 仅填写报道中**直接出现、可核对**的数据，不要推测或估算。
+2. 无明确依据的字段**必须省略**，不要填 0 或猜。
+3. **战损一律按增量**：只填本则报道中「本次/此次/今日/本轮」**新增**的伤亡或损毁数量。若报道只给「累计总数」「迄今共」「total so far」等，**不要填写**该字段（避免与库内已有累计值重复叠加）。
+4. **攻击地点**：提取双方遭袭的具体地点。美军/盟军基地被打击 → side=us；伊朗/亲伊设施被打击 → side=iran。name_keywords 用「中文名|英文名」便于匹配，可填多处。
 
-要求：
-- summary: 1-2句中文事实，≤80字
+字段说明：
+- summary: 1-2 句中文事实概括，≤80 字
 - category: deployment|alert|intel|diplomatic|other
 - severity: low|medium|high|critical
-- 战损（仅当新闻明确提及数字时填写，格式 us_XXX / iran_XXX）:
+- 战损（**仅填本则报道的新增增量**，如「此次 5 人丧生」「今日又损 2 架」）:
   us_personnel_killed, iran_personnel_killed, us_personnel_wounded, iran_personnel_wounded,
   us_civilian_killed, iran_civilian_killed, us_civilian_wounded, iran_civilian_wounded,
-  us_bases_destroyed, iran_bases_destroyed, us_bases_damaged, iran_bases_damaged.
-  重要：bases_* 仅指已确认损毁/受损的基地数量；"军事目标"/targets 等泛指不是基地，若报道只说"X个军事目标遭袭"而无具体基地名，不填写 bases_*
+  us_bases_destroyed, iran_bases_destroyed, us_bases_damaged, iran_bases_damaged,
   us_aircraft, iran_aircraft, us_warships, iran_warships, us_armor, iran_armor, us_vehicles, iran_vehicles,
   us_drones, iran_drones, us_missiles, iran_missiles, us_helicopters, iran_helicopters, us_submarines, iran_submarines,
-  us_tanks, iran_tanks, us_civilian_ships, iran_civilian_ships, us_airport_port, iran_airport_port
-- retaliation_sentiment: 0-100，仅当新闻涉及伊朗报复情绪时
-- wall_street_value: 0-100，仅当新闻涉及美股/市场反应时
-- key_location_updates: 当新闻提及具体基地/地点遭袭时，数组项 { "name_keywords": "asad|阿萨德|assad", "side": "us", "status": "attacked", "damage_level": 1-3 }
+  us_carriers, iran_carriers, us_civilian_ships, iran_civilian_ships, us_airport_port, iran_airport_port
+- retaliation_sentiment: 0-100，仅当报道涉及伊朗报复/反击情绪时
+- wall_street_value: 0-100，仅当报道涉及美股/市场时
+- key_location_updates: **双方攻击地点**。每项 {{ "name_keywords": "阿萨德|asad|al-asad", "side": "us或iran（被打击方）", "status": "attacked", "damage_level": 1-3 }}。美军基地例：阿萨德|asad、乌代德|udeid、埃尔比勒|erbil、因吉尔利克|incirlik。伊朗例：德黑兰|tehran、布什尔|bushehr、伊斯法罕|isfahan、阿巴斯|abbas、纳坦兹|natanz
 
-原文：{str(text)[:800]}
+原文：
+{raw}
 
 直接输出 JSON，不要解释："""
         r = requests.post(
@@ -48,7 +57,7 @@ def _call_ollama_extract(text: str, timeout: int = 10) -> Optional[Dict[str, Any
                 "model": OLLAMA_MODEL,
                 "messages": [{"role": "user", "content": prompt}],
                 "stream": False,
-                "options": {"num_predict": 256},
+                "options": {"num_predict": 384},
             },
             timeout=timeout,
         )
@@ -82,7 +91,7 @@ def extract_from_news(text: str, timestamp: Optional[str] = None) -> Dict[str, A
     # combat_losses 增量（仅数字字段）
     loss_us = {}
     loss_ir = {}
-    for k in ["personnel_killed", "personnel_wounded", "civilian_killed", "civilian_wounded", "bases_destroyed", "bases_damaged", "aircraft", "warships", "armor", "vehicles", "drones", "missiles", "helicopters", "submarines", "tanks", "civilian_ships", "airport_port"]:
+    for k in ["personnel_killed", "personnel_wounded", "civilian_killed", "civilian_wounded", "bases_destroyed", "bases_damaged", "aircraft", "warships", "armor", "vehicles", "drones", "missiles", "helicopters", "submarines", "carriers", "civilian_ships", "airport_port"]:
         uk = f"us_{k}"
         ik = f"iran_{k}"
         if uk in parsed and isinstance(parsed[uk], (int, float)):