第一次提交

2026-08-01 16:50:55 +08:00
commit c8f9e39039
68 changed files with 6016 additions and 0 deletions
@@ -0,0 +1,35 @@
+"""OCR 编排：Pix2Text 优先，失败/低置信/未部署时降级到多模态 VLM。"""
+import logging
+
+from app.config import settings
+from app.services.llm.client import LLMClient
+from app.services.ocr.base import OcrResult
+from app.services.ocr.pix2text_client import Pix2TextClient
+from app.services.ocr.vlm_engine import VlmOcrEngine
+
+logger = logging.getLogger(__name__)
+
+# Pix2Text 结果置信度低于此值则降级到 VLM
+_CONFIDENCE_THRESHOLD = 0.5
+
+
+async def run_ocr(image_bytes: bytes, mime: str, llm: LLMClient) -> OcrResult:
+    """返回识别结果，engine 字段标明最终使用的引擎。"""
+    # 1. 若配置了 Pix2Text 服务且健康，优先用
+    if settings.ocr_service_url:
+        client = Pix2TextClient(settings.ocr_service_url)
+        if await client.healthy():
+            try:
+                result = await client.recognize(image_bytes, mime)
+                if result.markdown.strip() and result.confidence >= _CONFIDENCE_THRESHOLD:
+                    return result
+                logger.info(
+                    "Pix2Text 置信度低(%.2f)或结果为空，降级 VLM", result.confidence
+                )
+            except Exception as e:  # noqa: BLE001
+                logger.warning("Pix2Text 识别失败，降级 VLM：%s", e)
+        else:
+            logger.info("Pix2Text 服务不健康，降级 VLM")
+
+    # 2. 降级 / 默认：多模态大模型
+    return await VlmOcrEngine(llm).recognize(image_bytes, mime)