Refactor: migrate pdf_parser.py to golang (#16323)

### What problem does this PR solve?

Http API based on onnx model.
pdf_parser.py to golang

### Type of change

- [x] Refactoring

deepdoc/server/endpoints/dla_endpoint.py

@@ -0,0 +1,43 @@
+"""DLA LitServe endpoint."""
+
+import logging
+
+import litserve as ls
+
+from deepdoc.server.adapters.dla_adapter import DLAAdapter
+
+logger = logging.getLogger(__name__)
+
+
+class DLAEndpoint(ls.LitAPI):
+    """Document Layout Analysis endpoint at /predict/dla."""
+
+    def __init__(self, model_dir: str, thr: float = 0.2):
+        super().__init__()
+        self.api_path = "/predict/dla"
+        self.model_dir = model_dir
+        self.thr = thr
+        self.adapter: DLAAdapter | None = None
+
+    def setup(self, device):
+        self.adapter = DLAAdapter(model_dir=self.model_dir, thr=self.thr)
+        self.adapter.load()
+        logger.info("DLA model loaded")
+
+    def decode_request(self, request):
+        # Handle both Starlette UploadFile (old) and FormData (Starlette >=1.3)
+        if hasattr(request, "file"):
+            data = request.file.read()
+        else:
+            data = request.get("request").file.read()
+        if not data:
+            raise ValueError("Empty request body")
+        if len(data) > 50 * 1024 * 1024:  # 50MB
+            raise ValueError("Image too large")
+        return data
+
+    def predict(self, image_data: bytes):
+        return self.adapter(image_data)
+
+    def encode_response(self, output):
+        return {"bboxes": output}

deepdoc/server/endpoints/ocr_endpoint.py

@@ -0,0 +1,67 @@
+"""OCR LitServe endpoint — detect + rec via operator form field."""
+
+import logging
+
+import litserve as ls
+
+from deepdoc.server.adapters.ocr_adapter import OCRAdapter
+
+logger = logging.getLogger(__name__)
+
+
+class OCREndpoint(ls.LitAPI):
+    """OCR endpoint at /predict/ocr.
+
+    Form field 'operator' (det or rec) selects the mode.
+    Form field 'request' carries the JPEG image bytes.
+    """
+
+    def __init__(self, model_dir: str):
+        super().__init__()
+        self.api_path = "/predict/ocr"
+        self.model_dir = model_dir
+        self.adapter: OCRAdapter | None = None
+
+    def setup(self, device):
+        self.adapter = OCRAdapter(model_dir=self.model_dir)
+        self.adapter.load()
+        logger.info("OCR model loaded")
+
+    def decode_request(self, request):
+        # Handle both old Starlette UploadFile and new Starlette FormData
+        if hasattr(request, "file"):
+            data = request.file.read()
+            # Try to read operator from the underlying request context
+            operator = getattr(self, "_request", None)
+            if operator is not None:
+                operator = operator.query_params.get("operator", "")
+            else:
+                operator = ""
+        else:
+            # FormData: get file and operator form fields
+            data = request.get("request").file.read()
+            op_val = request.get("operator")
+            operator = str(op_val) if op_val else ""
+
+        if not data:
+            raise ValueError("Empty request body")
+        if len(data) > 50 * 1024 * 1024:
+            raise ValueError("Image too large")
+
+        operator = operator.strip().lower()
+        if operator not in ("det", "rec"):
+            raise ValueError(
+                f"Invalid or missing operator '{operator}' (must be 'det' or 'rec')"
+            )
+
+        return operator, data
+
+    def predict(self, inputs: tuple):
+        operator, image_data = inputs
+        if operator == "det":
+            return self.adapter.detect(image_data)
+        else:
+            return self.adapter.recognize(image_data)
+
+    def encode_response(self, output):
+        return output

deepdoc/server/endpoints/tsr_endpoint.py

@@ -0,0 +1,43 @@
+"""TSR LitServe endpoint."""
+
+import logging
+
+import litserve as ls
+
+from deepdoc.server.adapters.tsr_adapter import TSRAdapter
+
+logger = logging.getLogger(__name__)
+
+
+class TSREndpoint(ls.LitAPI):
+    """Table Structure Recognition endpoint at /predict/tsr."""
+
+    def __init__(self, model_dir: str, thr: float = 0.2):
+        super().__init__()
+        self.api_path = "/predict/tsr"
+        self.model_dir = model_dir
+        self.thr = thr
+        self.adapter: TSRAdapter | None = None
+
+    def setup(self, device):
+        self.adapter = TSRAdapter(model_dir=self.model_dir, thr=self.thr)
+        self.adapter.load()
+        logger.info("TSR model loaded")
+
+    def decode_request(self, request):
+        # Handle both Starlette UploadFile (old) and FormData (Starlette >=1.3)
+        if hasattr(request, "file"):
+            data = request.file.read()
+        else:
+            data = request.get("request").file.read()
+        if not data:
+            raise ValueError("Empty request body")
+        if len(data) > 50 * 1024 * 1024:
+            raise ValueError("Image too large")
+        return data
+
+    def predict(self, image_data: bytes):
+        return self.adapter(image_data)
+
+    def encode_response(self, output):
+        return {"bboxes": output}