apps/worker/worker_app/tasks/vision/fast_path.py

@@ -21,9 +21,9 @@ logger = logging.getLogger(__name__)
 
 # 可通过环境变量调参（有默认值，无需配置即可跑）
 _IMG_WORKERS = int(os.environ.get("VISION_V2_IMG_WORKERS", "8"))
-_FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "6"))
-_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "5"))
-_OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "5"))
+_FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "8"))
+_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "8"))
+_OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "6"))
 _PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "45"))
 
 _FALLBACK_RESULT = {

apps/worker/worker_app/tasks/vision/vlm_fast_json.py

@@ -81,35 +81,79 @@ def call_fast_json(
 ) -> dict[str, Any] | None:
     """调用 lite VLM 返回结构化 dict；失败/非 JSON 返回 None。
 
-    注意：不做重试（外层竞速/降级逻辑负责），max_retries=0 由外层统一设置。
+    直接用 httpx 发最小 payload（关闭 thinking），不走 ai_client 包装：
+    - 关闭 thinking/推理链（reasoning_tokens 是延迟主因，单次要10-12s）
+    - 单次调用不重试（失败由外层走 pro 兜底）
+    - 温度=0.1 稳定输出 JSON
     """
     t0 = time.time()
+    import httpx
     try:
-        from packages.shared.ai_client import get_doubao_client
-
-        client = get_doubao_client()
-        if not client.is_available:
-            logger.warning("[vision.v2] doubao client 不可用，跳过 fast_json")
+        from packages.shared import get_shared_settings
+        settings = get_shared_settings()
+        api_key = settings.doubao_api_key
+        base_url = (settings.doubao_base_url or "https://ark.cn-beijing.volces.com/api/v3").rstrip("/")
+        if not api_key:
+            logger.warning("[vision.v2] doubao api_key 未配置，跳过 fast_json")
             return None
 
         use_model = model or DEFAULT_LITE_MODEL
-        # 强制不重试：lite 是快速路径，失败直接走外层 pro 兜底
-        _orig_retries = client.max_retries
-        client.max_retries = 0
+        url = f"{base_url}/chat/completions"
+        payload: dict[str, Any] = {
+            "model": use_model,
+            "messages": [
+                {"role": "system", "content": _FAST_SYSTEM},
+                {"role": "user", "content": [
+                    {"type": "image_url", "image_url": {"url": img_url}},
+                    {"type": "text", "text": _FAST_USER},
+                ]},
+            ],
+            "temperature": 0.1,
+            "max_tokens": max_tokens,
+            "stream": False,
+        }
+        # 关键：关闭 thinking（避免产生 reasoning_tokens 拖慢响应）
+        # 方舟/豆包 2.x 模型支持 thinking.type=disabled
         try:
-            raw = client.vision_completion(
-                messages=[
-                    {"role": "system", "content": _FAST_SYSTEM},
-                    {"role": "user", "content": _FAST_USER},
-                ],
-                images=[img_url],
-                temperature=0.1,
-                max_tokens=max_tokens,
-                timeout=timeout,
-                model=use_model,
-            )
-        finally:
-            client.max_retries = _orig_retries
+            payload["thinking"] = {"type": "disabled"}
+        except Exception:
+            pass
+        # 部分模型用 reasoning_effort 控制思考深度
+        payload["reasoning_effort"] = "low"
+
+        resp = httpx.post(
+            url,
+            headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+            json=payload,
+            timeout=timeout,
+        )
+        elapsed = time.time() - t0
+        if resp.status_code != 200:
+            logger.warning("[vision.v2] fast_json HTTP %d elapsed=%.1fs body=%s", resp.status_code, elapsed, resp.text[:200])
+            # 如果400说明不支持thinking参数，降级重试一次
+            if resp.status_code == 400 and "thinking" in resp.text.lower():
+                payload.pop("thinking", None)
+                payload.pop("reasoning_effort", None)
+                t1 = time.time()
+                resp = httpx.post(url,
+                    headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+                    json=payload, timeout=timeout,
+                )
+                elapsed = time.time() - t0
+                if resp.status_code != 200:
+                    logger.warning("[vision.v2] fast_json 降级后 HTTP %d elapsed=%.1fs", resp.status_code, elapsed)
+                    return None
+            else:
+                return None
+        data = resp.json()
+        raw = (data.get("choices") or [{}])[0].get("message", {}).get("content")
+        if raw is None:
+            logger.warning("[vision.v2] fast_json 返回 None elapsed=%.1fs", elapsed)
+            return None
+        usage = data.get("usage") or {}
+        logger.info("[vision.v2] fast_json 直连完成 model=%s elapsed=%.1fs in=%d out=%d reasoning=%d",
+            use_model, elapsed, usage.get("prompt_tokens",0), usage.get("completion_tokens",0),
+            usage.get("reasoning_tokens",0))
         elapsed = time.time() - t0
         if raw is None:
             logger.warning("[vision.v2] fast_json 返回 None elapsed=%.1fs model=%s", elapsed, use_model)