nv-ingest-api 2025.8.24.dev20250824__py3-none-any.whl → 2025.8.25.dev20250825__py3-none-any.whl

nv_ingest_api/internal/extract/image/chart_extractor.py

@@ -96,13 +96,23 @@ def _run_chart_inference(
         future_ocr_kwargs.update(
             model_name="paddle",
         )
-    else:
+    elif ocr_model_name == "scene_text":
         future_ocr_kwargs.update(
-            model_name="scene_text",
+            model_name=ocr_model_name,
             input_names=["input", "merge_levels"],
             dtypes=["FP32", "BYTES"],
             merge_level="paragraph",
         )
+    elif ocr_model_name == "scene_text_ensemble":
+        future_ocr_kwargs.update(
+            model_name=ocr_model_name,
+            input_names=["INPUT_IMAGE_URLS", "MERGE_LEVELS"],
+            output_names=["OUTPUT"],
+            dtypes=["BYTES", "BYTES"],
+            merge_level="paragraph",
+        )
+    else:
+        raise ValueError(f"Unknown OCR model name: {ocr_model_name}")
 
     with ThreadPoolExecutor(max_workers=2) as executor:
         future_yolox = executor.submit(yolox_client.infer, **future_yolox_kwargs)

nv_ingest_api/internal/extract/image/infographic_extractor.py

@@ -108,13 +108,24 @@ def _update_infographic_metadata(
         infer_kwargs.update(
             model_name="paddle",
         )
-    else:
+    elif ocr_model_name == "scene_text":
         infer_kwargs.update(
-            model_name="scene_text",
+            model_name=ocr_model_name,
             input_names=["input", "merge_levels"],
             dtypes=["FP32", "BYTES"],
             merge_level="paragraph",
         )
+    elif ocr_model_name == "scene_text_ensemble":
+        infer_kwargs.update(
+            model_name=ocr_model_name,
+            input_names=["INPUT_IMAGE_URLS", "MERGE_LEVELS"],
+            output_names=["OUTPUT"],
+            dtypes=["BYTES", "BYTES"],
+            merge_level="paragraph",
+        )
+    else:
+        raise ValueError(f"Unknown OCR model name: {ocr_model_name}")
+
     try:
         ocr_results = ocr_client.infer(data_ocr, **infer_kwargs)
     except Exception as e:

nv_ingest_api/internal/extract/image/table_extractor.py

@@ -96,13 +96,23 @@ def _run_inference(
         future_ocr_kwargs.update(
             model_name="paddle",
         )
-    else:
+    elif ocr_model_name == "scene_text":
         future_ocr_kwargs.update(
-            model_name="scene_text",
+            model_name=ocr_model_name,
             input_names=["input", "merge_levels"],
             dtypes=["FP32", "BYTES"],
             merge_level="word",
         )
+    elif ocr_model_name == "scene_text_ensemble":
+        future_ocr_kwargs.update(
+            model_name=ocr_model_name,
+            input_names=["INPUT_IMAGE_URLS", "MERGE_LEVELS"],
+            output_names=["OUTPUT"],
+            dtypes=["BYTES", "BYTES"],
+            merge_level="word",
+        )
+    else:
+        raise ValueError(f"Unknown OCR model name: {ocr_model_name}")
 
     with ThreadPoolExecutor(max_workers=2) as executor:
         future_ocr = executor.submit(ocr_client.infer, **future_ocr_kwargs)

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

@@ -26,8 +26,11 @@ from nv_ingest_api.internal.primitives.nim.model_interface.helpers import (
     preprocess_image_for_paddle,
 )
 from nv_ingest_api.util.image_processing.transforms import base64_to_numpy
+from nv_ingest_api.util.image_processing.transforms import numpy_to_base64
 
 DEFAULT_OCR_MODEL_NAME = "paddle"
+NEMORETRIEVER_OCR_EA_MODEL_NAME = "scene_text"
+NEMORETRIEVER_OCR_MODEL_NAME = "scene_text_ensemble"
 
 logger = logging.getLogger(__name__)
 
@@ -141,7 +144,7 @@ class OCRModelInterface(ModelInterface):
         images = data["image_arrays"]
         dims = data["image_dims"]
 
-        model_name = kwargs.get("model_name", "paddle")
+        model_name = kwargs.get("model_name", DEFAULT_OCR_MODEL_NAME)
         merge_level = kwargs.get("merge_level", "paragraph")
 
         if protocol == "grpc":
@@ -149,21 +152,33 @@ class OCRModelInterface(ModelInterface):
             processed: List[np.ndarray] = []
 
             max_length = max(max(img.shape[:2]) for img in images)
+            max_length = min(max_length, 65500)  # Maximum supported image dimension for JPEG is 65500 pixels.
 
             for img in images:
-                if model_name == "paddle":
+                if model_name == DEFAULT_OCR_MODEL_NAME:
                     arr, _dims = preprocess_image_for_paddle(img)
-                else:
+                elif model_name == NEMORETRIEVER_OCR_EA_MODEL_NAME:
                     arr, _dims = preprocess_image_for_ocr(
                         img,
                         target_height=max_length,
                         target_width=max_length,
                         pad_how="bottom_right",
                     )
+                elif model_name == NEMORETRIEVER_OCR_MODEL_NAME:
+                    arr = img
+                    _dims = {"new_width": img.shape[1], "new_height": img.shape[0]}
+                else:
+                    raise ValueError(f"Unknown model name: {model_name}")
 
                 dims.append(_dims)
-                arr = arr.astype(np.float32)
-                arr = np.expand_dims(arr, axis=0)  # => shape (1, H, W, C)
+
+                if model_name == NEMORETRIEVER_OCR_MODEL_NAME:
+                    arr = np.array([numpy_to_base64(arr, format="JPEG")], dtype=np.object_)
+                else:
+                    arr = arr.astype(np.float32)
+
+                arr = np.expand_dims(arr, axis=0)
+
                 processed.append(arr)
 
             batches = []
@@ -175,7 +190,7 @@ class OCRModelInterface(ModelInterface):
             ):
                 batched_input = np.concatenate(proc_chunk, axis=0)
 
-                if model_name == "paddle":
+                if model_name == DEFAULT_OCR_MODEL_NAME:
                     batches.append(batched_input)
                 else:
                     merge_levels = np.array([[merge_level] * len(batched_input)], dtype="object")
@@ -206,7 +221,7 @@ class OCRModelInterface(ModelInterface):
                 chunk_list(images, max_batch_size),
                 chunk_list(dims, max_batch_size),
             ):
-                if model_name == "paddle":
+                if model_name == DEFAULT_OCR_MODEL_NAME:
                     payload = {"input": input_chunk}
                 else:
                     payload = {
@@ -226,7 +241,7 @@ class OCRModelInterface(ModelInterface):
         response: Any,
         protocol: str,
         data: Optional[Dict[str, Any]] = None,
-        model_name: str = "paddle",
+        model_name: str = DEFAULT_OCR_MODEL_NAME,
         **kwargs: Any,
     ) -> Any:
         """
@@ -367,7 +382,7 @@ class OCRModelInterface(ModelInterface):
         self,
         response: np.ndarray,
         dimensions: List[Dict[str, Any]],
-        model_name: str = "paddle",
+        model_name: str = DEFAULT_OCR_MODEL_NAME,
     ) -> List[Tuple[str, str]]:
         """
         Parse a gRPC response for one or more images. The response can have two possible shapes:
@@ -402,12 +417,14 @@ class OCRModelInterface(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:
+            response = response.transpose((1, 0))
+
         # If we have shape (3,), convert to (3, 1)
         if response.ndim == 1 and response.shape == (3,):
             response = response.reshape(3, 1)
         elif response.ndim != 2 or response.shape[0] != 3:
             raise ValueError(f"Unexpected response shape: {response.shape}. Expecting (3,) or (3, n).")
-
         batch_size = response.shape[1]
         results: List[Tuple[str, str]] = []
 
@@ -425,11 +442,17 @@ class OCRModelInterface(ModelInterface):
             conf_scores = json.loads(confs_bytestr.decode("utf8"))
 
             # Some gRPC responses nest single-item lists; flatten them if needed
-            if isinstance(bounding_boxes, list) and len(bounding_boxes) == 1:
+            if (
+                (isinstance(bounding_boxes, list) and len(bounding_boxes) == 1 and isinstance(bounding_boxes[0], list))
+                and (
+                    isinstance(text_predictions, list)
+                    and len(text_predictions) == 1
+                    and isinstance(text_predictions[0], list)
+                )
+                and (isinstance(conf_scores, list) and len(conf_scores) == 1 and isinstance(conf_scores[0], list))
+            ):
                 bounding_boxes = bounding_boxes[0]
-            if isinstance(text_predictions, list) and len(text_predictions) == 1:
                 text_predictions = text_predictions[0]
-            if isinstance(conf_scores, list) and len(conf_scores) == 1:
                 conf_scores = conf_scores[0]
 
             # 4) Postprocess
@@ -439,7 +462,7 @@ class OCRModelInterface(ModelInterface):
                 conf_scores,
                 dimensions,
                 img_index=i,
-                scale_coordinates=True if model_name == "paddle" else False,
+                scale_coordinates=False if model_name == NEMORETRIEVER_OCR_EA_MODEL_NAME else True,
             )
 
             results.append([bounding_boxes, text_predictions, conf_scores])

{nv_ingest_api-2025.8.24.dev20250824.dist-info → nv_ingest_api-2025.8.25.dev20250825.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nv-ingest-api
-Version: 2025.8.24.dev20250824
+Version: 2025.8.25.dev20250825
 Summary: Python module with core document ingestion functions.
 Author-email: Jeremy Dyer <jdyer@nvidia.com>
 License:                                  Apache License

{nv_ingest_api-2025.8.24.dev20250824.dist-info → nv_ingest_api-2025.8.25.dev20250825.dist-info}/RECORD

@@ -20,10 +20,10 @@ nv_ingest_api/internal/extract/docx/engines/docxreader_helpers/docxreader.py,sha
 nv_ingest_api/internal/extract/html/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/internal/extract/html/html_extractor.py,sha256=I9oWfj6_As4898GDDh0zsSuKxO3lBsvyYzhvUotjzJI,3282
 nv_ingest_api/internal/extract/image/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
-nv_ingest_api/internal/extract/image/chart_extractor.py,sha256=gk-O-9wjZBoaLVE_6Erb4gMwsSFk4UtPQ2QLpMCW4H4,13212
+nv_ingest_api/internal/extract/image/chart_extractor.py,sha256=HTu0mOIYMpK5JAAiyvlIHgUtmjQV44Gv8dE5hYIqeQE,13633
 nv_ingest_api/internal/extract/image/image_extractor.py,sha256=gBKjlx28hA_e-dupatu46YQgOHJ0DLpAWxREiLaZLyo,9039
-nv_ingest_api/internal/extract/image/infographic_extractor.py,sha256=i7zt_ow1gytU4hK2JCRg7T1wlbokaeuUpXX69LIQkzY,9687
-nv_ingest_api/internal/extract/image/table_extractor.py,sha256=O0m3N2Tz9W6X7TBI4o-rbBXc8dFOf9zSZq1v9qC1U4M,13780
+nv_ingest_api/internal/extract/image/infographic_extractor.py,sha256=G5sRnyJ-8ToBbD0_7W6Vemq4a5SBNLtzhZKpuR26mlU,10104
+nv_ingest_api/internal/extract/image/table_extractor.py,sha256=yjSehCTV43a35I_JrVNkgi7yV6RTAEvTeB3kGtM9ZTs,14196
 nv_ingest_api/internal/extract/image/image_helpers/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/internal/extract/image/image_helpers/common.py,sha256=80jRhGzisHvQ9Ky3MKUMM7soKUmvZ5LqRVzwNYjgdPY,14988
 nv_ingest_api/internal/extract/pdf/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
@@ -58,7 +58,7 @@ nv_ingest_api/internal/primitives/nim/model_interface/decorators.py,sha256=qwubk
 nv_ingest_api/internal/primitives/nim/model_interface/deplot.py,sha256=TvKdk6PTuI1WNhRmNNrvygaI_DIutkJkDL-XdtLZQac,10787
 nv_ingest_api/internal/primitives/nim/model_interface/helpers.py,sha256=iyGxAr4tG2UZ7LtXXoWO_kF-KsObhPrmZ46Nl0Mi-Ag,11592
 nv_ingest_api/internal/primitives/nim/model_interface/nemoretriever_parse.py,sha256=WysjDZeegclO3mZgVcGOwzWbr8wSI4pWRiYD4iC2EXo,7098
-nv_ingest_api/internal/primitives/nim/model_interface/ocr.py,sha256=Vhim3py_rc5jA0BoKubwfekEqOwxUUePzcmc59pRuOk,21458
+nv_ingest_api/internal/primitives/nim/model_interface/ocr.py,sha256=0Xfuf5_-7LoWnqzZlsJFI53ztneB7Rs-PHZQzDgR0mo,22679
 nv_ingest_api/internal/primitives/nim/model_interface/parakeet.py,sha256=5PqD2JuHY2rwd-6SSB4axr2Dd79vm95sAEkcmI3U7ME,12977
 nv_ingest_api/internal/primitives/nim/model_interface/text_embedding.py,sha256=lFhppNqrq5X_fzbCWKphvZQMzaJd3gHrkWsyJORzFrU,5010
 nv_ingest_api/internal/primitives/nim/model_interface/vlm.py,sha256=qJ382PU1ZrIM-SR3cqIhtY_W2rmHec2HIa2aUB2SvaU,6031
@@ -162,8 +162,8 @@ nv_ingest_api/util/string_processing/configuration.py,sha256=2HS08msccuPCT0fn_jf
 nv_ingest_api/util/string_processing/yaml.py,sha256=6SW2O6wbXRhGbhETMbtXjYCZn53HeCNOP6a96AaxlHs,1454
 nv_ingest_api/util/system/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nv_ingest_api/util/system/hardware_info.py,sha256=1UFM8XE6M3pgQcpbVsCsqDQ7Dj-zzptL-XRE-DEu9UA,27213
-nv_ingest_api-2025.8.24.dev20250824.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-nv_ingest_api-2025.8.24.dev20250824.dist-info/METADATA,sha256=efa0KmmDvdvacCuWPIHCoYGvujzWjw-LIdWdnhSW3Cw,13947
-nv_ingest_api-2025.8.24.dev20250824.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-nv_ingest_api-2025.8.24.dev20250824.dist-info/top_level.txt,sha256=abjYMlTJGoG5tOdfIB-IWvLyKclw6HLaRSc8MxX4X6I,14
-nv_ingest_api-2025.8.24.dev20250824.dist-info/RECORD,,
+nv_ingest_api-2025.8.25.dev20250825.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+nv_ingest_api-2025.8.25.dev20250825.dist-info/METADATA,sha256=IBhn7pRL6SlKTG59w68Mo4Gets_IBm_rBOTyDf2aZXU,13947
+nv_ingest_api-2025.8.25.dev20250825.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+nv_ingest_api-2025.8.25.dev20250825.dist-info/top_level.txt,sha256=abjYMlTJGoG5tOdfIB-IWvLyKclw6HLaRSc8MxX4X6I,14
+nv_ingest_api-2025.8.25.dev20250825.dist-info/RECORD,,