nv-ingest-api 2025.10.15.dev20251015__tar.gz → 2025.10.17.dev20251017__tar.gz

{nv_ingest_api-2025.10.15.dev20251015/src/nv_ingest_api.egg-info → nv_ingest_api-2025.10.17.dev20251017}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nv-ingest-api
-Version: 2025.10.15.dev20251015
+Version: 2025.10.17.dev20251017
 Summary: Python module with core document ingestion functions.
 Author-email: Jeremy Dyer <jdyer@nvidia.com>
 License:                                  Apache License

{nv_ingest_api-2025.10.15.dev20251015 → nv_ingest_api-2025.10.17.dev20251017}/src/nv_ingest_api/internal/primitives/nim/model_interface/ocr.py

@@ -758,8 +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-2025.10.15.dev20251015 → nv_ingest_api-2025.10.17.dev20251017/src/nv_ingest_api.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nv-ingest-api
-Version: 2025.10.15.dev20251015
+Version: 2025.10.17.dev20251017
 Summary: Python module with core document ingestion functions.
 Author-email: Jeremy Dyer <jdyer@nvidia.com>
 License:                                  Apache License

{nv_ingest_api-2025.10.15.dev20251015 → nv_ingest_api-2025.10.17.dev20251017}/src/udfs/llm_summarizer_udf.py

@@ -2,8 +2,9 @@
 """
 LLM Content Summarizer UDF for NV-Ingest Pipeline
 
-This UDF uses an LLM to generate concise summaries of text content chunks. These summaries are added to the metadata
-for enhanced downstream processing and search capabilities.
+Generates document summaries using NVIDIA-hosted LLMs. This production UDF demonstrates how to extract the pipeline
+payload, run custom code (summarization), and inject results into the metadata for downstream usecases (such as
+retrieval).
 
 These variables can be set in the environment before running the pipeline. These can be treated as kwargs.
 - NVIDIA_API_KEY: API key for NVIDIA NIM endpoints (required)
@@ -14,16 +15,14 @@ These variables can be set in the environment before running the pipeline. These
 - MAX_CONTENT_LENGTH: Maximum content length to send to API (default: 12000)
 TODO: Implement this
 - NUM_CHUNKS: (Optional) Number of first and last pages to summarize. default=1
+
+More info can be found in `examples/udfs/README.md`
 """
 
 import logging
 import os
 import time
 
-# REMOVE BEFORE MERGING
-# import yaml
-# from pathlib import Path
-
 
 logger = logging.getLogger(__name__)
 
@@ -85,6 +84,7 @@ def content_summarizer(control_message: "IngestControlMessage") -> "IngestContro
     # According to docs/docs/extraction/user_defined_functions.md#understanding-the-dataframe-payload
     # the rows are not necessarily pages. they are chunks of data extracted from the document. in order to select
     # pages, it must require parsing the payload to see which chunks correspond to which pages
+    original_df = df.copy()
     if len(df) > 1:
         # TODO: add feature to select N first and last chunks
         df = df.iloc[[0, -1]]
@@ -108,10 +108,11 @@ def content_summarizer(control_message: "IngestControlMessage") -> "IngestContro
     if not stats["failed"]:
         stats["tokens"] = _estimate_tokens(content)
         logger.info("Summarized %d tokens in %f seconds using %s", stats["tokens"], stats["duration"], model_name)
-        _store_summary(df, summary, model_name)
+        _store_summary(original_df, summary, model_name)
 
         # Update the control message with modified DataFrame
-        control_message.payload(df)
+        control_message.payload(original_df)
+
     else:
         logger.warning("%s failed to summarize content", model_name)