nv-ingest-api 2025.10.4.dev20251004__py3-none-any.whl → 2025.11.2.dev20251102__py3-none-any.whl

nv_ingest_api/internal/extract/image/chart_extractor.py

@@ -97,7 +97,7 @@ def _run_chart_inference(
             model_name="paddle",
             max_batch_size=1 if ocr_client.protocol == "grpc" else 2,
         )
-    elif ocr_model_name == "scene_text_ensemble":
+    elif ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"}:
         future_ocr_kwargs.update(
             model_name=ocr_model_name,
             input_names=["INPUT_IMAGE_URLS", "MERGE_LEVELS"],
@@ -237,7 +237,9 @@ def _create_ocr_client(
     auth_token: str,
 ) -> NimClient:
     ocr_model_interface = (
-        NemoRetrieverOCRModelInterface() if ocr_model_name == "scene_text_ensemble" else PaddleOCRModelInterface()
+        NemoRetrieverOCRModelInterface()
+        if ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"}
+        else PaddleOCRModelInterface()
     )
 
     ocr_client = create_inference_client(
@@ -245,7 +247,9 @@ def _create_ocr_client(
         model_interface=ocr_model_interface,
         auth_token=auth_token,
         infer_protocol=ocr_protocol,
-        enable_dynamic_batching=(True if ocr_model_name == "scene_text_ensemble" else False),
+        enable_dynamic_batching=(
+            True if ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"} else False
+        ),
         dynamic_batch_memory_budget_mb=32,
     )
 

nv_ingest_api/internal/extract/image/infographic_extractor.py

@@ -107,7 +107,7 @@ def _update_infographic_metadata(
             model_name="paddle",
             max_batch_size=1 if ocr_client.protocol == "grpc" else 2,
         )
-    elif ocr_model_name == "scene_text_ensemble":
+    elif ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"}:
         infer_kwargs.update(
             model_name=ocr_model_name,
             input_names=["INPUT_IMAGE_URLS", "MERGE_LEVELS"],
@@ -152,7 +152,9 @@ def _create_ocr_client(
     auth_token: str,
 ) -> NimClient:
     ocr_model_interface = (
-        NemoRetrieverOCRModelInterface() if ocr_model_name == "scene_text_ensemble" else PaddleOCRModelInterface()
+        NemoRetrieverOCRModelInterface()
+        if ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"}
+        else PaddleOCRModelInterface()
     )
 
     ocr_client = create_inference_client(
@@ -160,7 +162,9 @@ def _create_ocr_client(
         model_interface=ocr_model_interface,
         auth_token=auth_token,
         infer_protocol=ocr_protocol,
-        enable_dynamic_batching=(True if ocr_model_name == "scene_text_ensemble" else False),
+        enable_dynamic_batching=(
+            True if ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"} else False
+        ),
         dynamic_batch_memory_budget_mb=32,
     )
 

nv_ingest_api/internal/extract/image/table_extractor.py

@@ -99,7 +99,7 @@ def _run_inference(
             model_name="paddle",
             max_batch_size=1 if ocr_client.protocol == "grpc" else 2,
         )
-    elif ocr_model_name == "scene_text_ensemble":
+    elif ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"}:
         future_ocr_kwargs.update(
             model_name=ocr_model_name,
             input_names=["INPUT_IMAGE_URLS", "MERGE_LEVELS"],
@@ -246,7 +246,9 @@ def _create_ocr_client(
     auth_token: str,
 ) -> NimClient:
     ocr_model_interface = (
-        NemoRetrieverOCRModelInterface() if ocr_model_name == "scene_text_ensemble" else PaddleOCRModelInterface()
+        NemoRetrieverOCRModelInterface()
+        if ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"}
+        else PaddleOCRModelInterface()
     )
 
     ocr_client = create_inference_client(
@@ -254,7 +256,9 @@ def _create_ocr_client(
         model_interface=ocr_model_interface,
         auth_token=auth_token,
         infer_protocol=ocr_protocol,
-        enable_dynamic_batching=(True if ocr_model_name == "scene_text_ensemble" else False),
+        enable_dynamic_batching=(
+            True if ocr_model_name in {"scene_text_ensemble", "scene_text_wrapper", "scene_text_python"} else False
+        ),
         dynamic_batch_memory_budget_mb=32,
     )
 

nv_ingest_api/internal/extract/pdf/engines/pdfium.py

@@ -332,6 +332,7 @@ def _extract_page_elements(
 
         # Process each extracted element based on extraction flags
         for page_idx, page_element in page_element_results:
+            page_reading_index = page_idx + 1
             # Skip elements that shouldn't be extracted based on flags
             if (not extract_tables) and (page_element.type_string == "table"):
                 continue
@@ -347,7 +348,7 @@ def _extract_page_elements(
             # Construct metadata for the page element
             page_element_meta = construct_page_element_metadata(
                 page_element,
-                page_idx,
+                page_reading_index,
                 page_count,
                 source_metadata,
                 base_unified_metadata,
@@ -473,6 +474,7 @@ def pdfium_extractor(
         for page_idx in range(page_count):
             page = doc.get_page(page_idx)
             page_width, page_height = page.get_size()
+            page_reading_index = page_idx + 1
 
             # Text extraction
             if extract_text:
@@ -481,7 +483,7 @@ def pdfium_extractor(
                     text_meta = construct_text_metadata(
                         [page_text],
                         pdf_metadata.keywords,
-                        page_idx,
+                        page_reading_index,
                         -1,
                         -1,
                         -1,
@@ -499,7 +501,7 @@ def pdfium_extractor(
                 image_data = _extract_page_images(
                     extract_images_method,
                     page,
-                    page_idx,
+                    page_reading_index,
                     page_width,
                     page_height,
                     page_count,
@@ -518,7 +520,7 @@ def pdfium_extractor(
                     base64_image, _ = scale_image_to_encoding_size(base64_image, max_base64_size=2**24 - 1)
                 image_meta = construct_image_metadata_from_base64(
                     base64_image,
-                    page_idx,
+                    page_reading_index,
                     page_count,
                     source_metadata,
                     base_unified_metadata,

nv_ingest_api/internal/primitives/nim/model_interface/ocr.py

@@ -21,7 +21,10 @@ from nv_ingest_api.internal.primitives.nim.model_interface.helpers import prepro
 from nv_ingest_api.util.image_processing.transforms import base64_to_numpy
 
 DEFAULT_OCR_MODEL_NAME = "paddle"
-NEMORETRIEVER_OCR_MODEL_NAME = "scene_text_ensemble"
+NEMORETRIEVER_OCR_MODEL_NAME = "scene_text_wrapper"
+NEMORETRIEVER_OCR_ENSEMBLE_MODEL_NAME = "scene_text_ensemble"
+NEMORETRIEVER_OCR_BLS_MODEL_NAME = "scene_text_python"
+
 
 logger = logging.getLogger(__name__)
 
@@ -231,7 +234,11 @@ class OCRModelInterfaceBase(ModelInterface):
         if not isinstance(response, np.ndarray):
             raise ValueError("Unexpected response format: response is not a NumPy array.")
 
-        if model_name == NEMORETRIEVER_OCR_MODEL_NAME:
+        if model_name in [
+            NEMORETRIEVER_OCR_MODEL_NAME,
+            NEMORETRIEVER_OCR_ENSEMBLE_MODEL_NAME,
+            NEMORETRIEVER_OCR_BLS_MODEL_NAME,
+        ]:
             response = response.transpose((1, 0))
 
         # If we have shape (3,), convert to (3, 1)
@@ -751,8 +758,8 @@ def get_ocr_model_name(ocr_grpc_endpoint=None, default_model_name=DEFAULT_OCR_MO
     if ocr_model_name is not None:
         return ocr_model_name
 
-    # 2. If no gRPC endpoint is provided, fall back to the default immediately.
-    if not ocr_grpc_endpoint:
+    # 2. If no gRPC endpoint is provided or the endpoint is a NVCF endpoint, fall back to the default immediately.
+    if (not ocr_grpc_endpoint) or ("grpc.nvcf.nvidia.com" in ocr_grpc_endpoint):
         logger.debug(f"No OCR gRPC endpoint provided. Falling back to default model name '{default_model_name}'.")
         return default_model_name
 

nv_ingest_api/internal/primitives/nim/model_interface/parakeet.py

@@ -355,6 +355,10 @@ def create_audio_inference_client(
     if (infer_protocol is None) and (grpc_endpoint and grpc_endpoint.strip()):
         infer_protocol = "grpc"
 
+    # Normalize protocol to lowercase for case-insensitive comparison
+    if infer_protocol:
+        infer_protocol = infer_protocol.lower()
+
     if infer_protocol == "http":
         raise ValueError("`http` endpoints are not supported for audio. Use `grpc`.")
 

nv_ingest_api/internal/primitives/nim/nim_client.py

@@ -5,6 +5,7 @@
 import hashlib
 import json
 import logging
+import re
 import threading
 import time
 import queue
@@ -24,6 +25,12 @@ from nv_ingest_api.util.string_processing import generate_url
 
 logger = logging.getLogger(__name__)
 
+# Regex pattern to detect CUDA-related errors in Triton gRPC responses
+CUDA_ERROR_REGEX = re.compile(
+    r"(illegal memory access|invalid argument|failed to (copy|load|perform) .*: .*|TritonModelException: failed to copy data: .*)",  # noqa: E501
+    re.IGNORECASE,
+)
+
 # A simple structure to hold a request's data and its Future for the result
 InferenceRequest = namedtuple("InferenceRequest", ["data", "future", "model_name", "dims", "kwargs"])
 
@@ -40,7 +47,7 @@ class NimClient:
         endpoints: Tuple[str, str],
         auth_token: Optional[str] = None,
         timeout: float = 120.0,
-        max_retries: int = 5,
+        max_retries: int = 10,
         max_429_retries: int = 5,
         enable_dynamic_batching: bool = False,
         dynamic_batch_timeout: float = 0.1,  # 100 milliseconds
@@ -60,11 +67,11 @@ class NimClient:
         auth_token : str, optional
             Authorization token for HTTP requests (default: None).
         timeout : float, optional
-            Timeout for HTTP requests in seconds (default: 30.0).
+            Timeout for HTTP requests in seconds (default: 120.0).
         max_retries : int, optional
-            The maximum number of retries for non-429 server-side errors (default: 5).
+            The maximum number of retries for non-429 server-side errors (default: 10).
         max_429_retries : int, optional
-            The maximum number of retries specifically for 429 errors (default: 10).
+            The maximum number of retries specifically for 429 errors (default: 5).
 
         Raises
         ------
@@ -121,9 +128,6 @@ class NimClient:
         if model_name == "yolox_ensemble":
             model_name = "yolox"
 
-        if model_name == "scene_text_ensemble":
-            model_name = "scene_text_pre"
-
         if model_name in self._max_batch_sizes:
             return self._max_batch_sizes[model_name]
 
@@ -326,16 +330,101 @@ class NimClient:
 
         outputs = [grpcclient.InferRequestedOutput(output_name) for output_name in output_names]
 
-        response = self.client.infer(
-            model_name=model_name, parameters=parameters, inputs=input_tensors, outputs=outputs
-        )
+        base_delay = 2.0
+        attempt = 0
+        retries_429 = 0
+        max_grpc_retries = self.max_429_retries
 
-        logger.debug(f"gRPC inference response: {response}")
+        while attempt < self.max_retries:
+            try:
+                response = self.client.infer(
+                    model_name=model_name, parameters=parameters, inputs=input_tensors, outputs=outputs
+                )
 
-        if len(outputs) == 1:
-            return response.as_numpy(outputs[0].name())
-        else:
-            return [response.as_numpy(output.name()) for output in outputs]
+                logger.debug(f"gRPC inference response: {response}")
+
+                if len(outputs) == 1:
+                    return response.as_numpy(outputs[0].name())
+                else:
+                    return [response.as_numpy(output.name()) for output in outputs]
+
+            except grpcclient.InferenceServerException as e:
+                status = str(e.status())
+                message = e.message()
+
+                # Handle CUDA memory errors
+                if status == "StatusCode.INTERNAL":
+                    if CUDA_ERROR_REGEX.search(message):
+                        logger.warning(
+                            f"Received gRPC INTERNAL error with CUDA-related message for model '{model_name}'. "
+                            f"Attempt {attempt + 1} of {self.max_retries}. Message (truncated): {message[:500]}"
+                        )
+                        if attempt >= self.max_retries - 1:
+                            logger.error(f"Max retries exceeded for CUDA errors on model '{model_name}'.")
+                            raise e
+                        # Try to reload models before retrying
+                        model_reload_succeeded = reload_models(client=self.client, client_timeout=self.timeout)
+                        if not model_reload_succeeded:
+                            logger.error(f"Failed to reload models for model '{model_name}'.")
+                    else:
+                        logger.warning(
+                            f"Received gRPC INTERNAL error for model '{model_name}'. "
+                            f"Attempt {attempt + 1} of {self.max_retries}. Message (truncated): {message[:500]}"
+                        )
+                        if attempt >= self.max_retries - 1:
+                            logger.error(f"Max retries exceeded for INTERNAL error on model '{model_name}'.")
+                            raise e
+
+                    # Common retry logic for both CUDA and non-CUDA INTERNAL errors
+                    backoff_time = base_delay * (2**attempt)
+                    time.sleep(backoff_time)
+                    attempt += 1
+                    continue
+
+                # Handle errors that can occur after model reload (NOT_FOUND, model not loaded)
+                if status == "StatusCode.NOT_FOUND":
+                    logger.warning(
+                        f"Received gRPC {status} error for model '{model_name}'. "
+                        f"Attempt {attempt + 1} of {self.max_retries}. Message: {message[:500]}"
+                    )
+                    if attempt >= self.max_retries - 1:
+                        logger.error(f"Max retries exceeded for model not found errors on model '{model_name}'.")
+                        raise e
+
+                    # Retry with exponential backoff WITHOUT reloading
+                    backoff_time = base_delay * (2**attempt)
+                    logger.info(
+                        f"Retrying after {backoff_time}s backoff for model not found error on model '{model_name}'."
+                    )
+                    time.sleep(backoff_time)
+                    attempt += 1
+                    continue
+
+                if status == "StatusCode.UNAVAILABLE" and "Exceeds maximum queue size".lower() in message.lower():
+                    retries_429 += 1
+                    logger.warning(
+                        f"Received gRPC {status} for model '{model_name}'. "
+                        f"Attempt {retries_429} of {max_grpc_retries}."
+                    )
+                    if retries_429 >= max_grpc_retries:
+                        logger.error(f"Max retries for gRPC {status} exceeded for model '{model_name}'.")
+                        raise
+
+                    backoff_time = base_delay * (2**retries_429)
+                    time.sleep(backoff_time)
+                    continue
+
+                # For other server-side errors (e.g., INVALID_ARGUMENT, etc.),
+                # fail fast as retrying will not help
+                logger.error(
+                    f"Received non-retryable gRPC error {status} from Triton for model '{model_name}': {message}"
+                )
+                raise
+
+            except Exception as e:
+                # Catch any other unexpected exceptions (e.g., network issues not caught by Triton client)
+                logger.error(f"An unexpected error occurred during gRPC inference for model '{model_name}': {e}")
+                raise
 
     def _http_infer(self, formatted_input: dict) -> dict:
         """
@@ -648,3 +737,57 @@ class NimClientManager:
 def get_nim_client_manager(*args, **kwargs) -> NimClientManager:
     """Returns the singleton instance of the NimClientManager."""
     return NimClientManager(*args, **kwargs)
+
+
+def reload_models(client: grpcclient.InferenceServerClient, exclude: list[str] = [], client_timeout: int = 120) -> bool:
+    """
+    Reloads all models in the Triton server except for the models in the exclude list.
+
+    Parameters
+    ----------
+    client : grpcclient.InferenceServerClient
+        The gRPC client connected to the Triton server.
+    exclude : list[str], optional
+        A list of model names to exclude from reloading.
+    client_timeout : int, optional
+        Timeout for client operations in seconds (default: 120).
+
+    Returns
+    -------
+    bool
+        True if all models were successfully reloaded, False otherwise.
+    """
+    model_index = client.get_model_repository_index()
+    exclude = set(exclude)
+    names = [m.name for m in model_index.models if m.name not in exclude]
+
+    logger.info(f"Reloading {len(names)} model(s): {', '.join(names) if names else '(none)'}")
+
+    # 1) Unload
+    for name in names:
+        try:
+            client.unload_model(name)
+        except grpcclient.InferenceServerException as e:
+            msg = e.message()
+            if "explicit model load / unload" in msg.lower():
+                status = e.status()
+                logger.warning(
+                    f"[SKIP Model Reload] Explicit model control disabled; cannot unload '{name}'. Status: {status}."
+                )
+                return False
+            logger.error(f"[ERROR] Failed to unload '{name}': {msg}")
+            return False
+
+    # 2) Load
+    for name in names:
+        client.load_model(name)
+
+    # 3) Readiness check
+    for name in names:
+        ready = client.is_model_ready(model_name=name, client_timeout=client_timeout)
+        if not ready:
+            logger.warning(f"[Warning] Triton Not ready: {name}")
+            return False
+
+    logger.info("✅ Reload of models complete.")
+    return True

nv_ingest_api/internal/schemas/extract/extract_audio_schema.py

@@ -10,10 +10,12 @@ from typing import Tuple
 from pydantic import BaseModel, Field
 from pydantic import root_validator
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class AudioConfigSchema(BaseModel):
+class AudioConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for audio extraction endpoints and options.
 
@@ -87,13 +89,13 @@ class AudioConfigSchema(BaseModel):
 
         values[endpoint_name] = (grpc_service, http_service)
 
+        # Auto-infer protocol from endpoints if not specified
         protocol_name = "audio_infer_protocol"
         protocol_value = values.get(protocol_name)
 
         if not protocol_value:
             protocol_value = "http" if http_service else "grpc" if grpc_service else ""
 
-        protocol_value = protocol_value.lower()
         values[protocol_name] = protocol_value
 
         return values

nv_ingest_api/internal/schemas/extract/extract_chart_schema.py

@@ -8,10 +8,12 @@ from typing import Tuple
 
 from pydantic import field_validator, model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class ChartExtractorConfigSchema(BaseModel):
+class ChartExtractorConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for chart extraction service endpoints and options.
 
@@ -96,6 +98,13 @@ class ChartExtractorConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
+            protocol_name = endpoint_name.replace("_endpoints", "_infer_protocol")
+            protocol_value = values.get(protocol_name)
+            if not protocol_value:
+                protocol_value = "http" if http_service else "grpc" if grpc_service else ""
+            values[protocol_name] = protocol_value
+
         return values
 
     model_config = ConfigDict(extra="forbid")

nv_ingest_api/internal/schemas/extract/extract_docx_schema.py

@@ -9,10 +9,12 @@ from typing import Tuple
 
 from pydantic import model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class DocxConfigSchema(BaseModel):
+class DocxConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for docx extraction endpoints and options.
 
@@ -85,11 +87,11 @@ class DocxConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
             protocol_name = f"{model_name}_infer_protocol"
             protocol_value = values.get(protocol_name)
             if not protocol_value:
                 protocol_value = "http" if http_service else "grpc" if grpc_service else ""
-            protocol_value = protocol_value.lower()
             values[protocol_name] = protocol_value
 
         return values

nv_ingest_api/internal/schemas/extract/extract_image_schema.py

@@ -9,10 +9,12 @@ from typing import Tuple
 
 from pydantic import model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class ImageConfigSchema(BaseModel):
+class ImageConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for image extraction endpoints and options.
 
@@ -85,11 +87,11 @@ class ImageConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
             protocol_name = f"{model_name}_infer_protocol"
             protocol_value = values.get(protocol_name)
             if not protocol_value:
                 protocol_value = "http" if http_service else "grpc" if grpc_service else ""
-            protocol_value = protocol_value.lower()
             values[protocol_name] = protocol_value
 
         return values

nv_ingest_api/internal/schemas/extract/extract_infographic_schema.py

@@ -8,10 +8,12 @@ from typing import Tuple
 
 from pydantic import field_validator, model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class InfographicExtractorConfigSchema(BaseModel):
+class InfographicExtractorConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for infographic extraction service endpoints and options.
 
@@ -89,6 +91,13 @@ class InfographicExtractorConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
+            protocol_name = endpoint_name.replace("_endpoints", "_infer_protocol")
+            protocol_value = values.get(protocol_name)
+            if not protocol_value:
+                protocol_value = "http" if http_service else "grpc" if grpc_service else ""
+            values[protocol_name] = protocol_value
+
         return values
 
     model_config = ConfigDict(extra="forbid")

nv_ingest_api/internal/schemas/extract/extract_pdf_schema.py

@@ -9,10 +9,12 @@ from typing import Tuple
 
 from pydantic import model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class PDFiumConfigSchema(BaseModel):
+class PDFiumConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for PDFium endpoints and options.
 
@@ -82,11 +84,11 @@ class PDFiumConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
             protocol_name = f"{model_name}_infer_protocol"
             protocol_value = values.get(protocol_name)
             if not protocol_value:
                 protocol_value = "http" if http_service else "grpc" if grpc_service else ""
-            protocol_value = protocol_value.lower()
             values[protocol_name] = protocol_value
 
         return values
@@ -94,7 +96,7 @@ class PDFiumConfigSchema(BaseModel):
     model_config = ConfigDict(extra="forbid")
 
 
-class NemoRetrieverParseConfigSchema(BaseModel):
+class NemoRetrieverParseConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for NemoRetrieverParse endpoints and options.
 
@@ -170,11 +172,11 @@ class NemoRetrieverParseConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
             protocol_name = f"{model_name}_infer_protocol"
             protocol_value = values.get(protocol_name)
             if not protocol_value:
                 protocol_value = "http" if http_service else "grpc" if grpc_service else ""
-            protocol_value = protocol_value.lower()
             values[protocol_name] = protocol_value
 
         return values

nv_ingest_api/internal/schemas/extract/extract_pptx_schema.py

@@ -9,10 +9,12 @@ from typing import Tuple
 
 from pydantic import model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
+
 logger = logging.getLogger(__name__)
 
 
-class PPTXConfigSchema(BaseModel):
+class PPTXConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for docx extraction endpoints and options.
 
@@ -85,11 +87,11 @@ class PPTXConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
             protocol_name = f"{model_name}_infer_protocol"
             protocol_value = values.get(protocol_name)
             if not protocol_value:
                 protocol_value = "http" if http_service else "grpc" if grpc_service else ""
-            protocol_value = protocol_value.lower()
             values[protocol_name] = protocol_value
 
         return values

nv_ingest_api/internal/schemas/extract/extract_table_schema.py

@@ -9,11 +9,12 @@ from typing import Tuple
 
 from pydantic import field_validator, model_validator, ConfigDict, BaseModel, Field
 
+from nv_ingest_api.internal.schemas.mixins import LowercaseProtocolMixin
 
 logger = logging.getLogger(__name__)
 
 
-class TableExtractorConfigSchema(BaseModel):
+class TableExtractorConfigSchema(LowercaseProtocolMixin):
     """
     Configuration schema for the table extraction stage settings.
 
@@ -91,6 +92,13 @@ class TableExtractorConfigSchema(BaseModel):
 
             values[endpoint_name] = (grpc_service, http_service)
 
+            # Auto-infer protocol from endpoints if not specified
+            protocol_name = endpoint_name.replace("_endpoints", "_infer_protocol")
+            protocol_value = values.get(protocol_name)
+            if not protocol_value:
+                protocol_value = "http" if http_service else "grpc" if grpc_service else ""
+            values[protocol_name] = protocol_value
+
         return values
 
     model_config = ConfigDict(extra="forbid")