Integrate NVIDIA Retriever Extraction with Langflow

The NVIDIA Retriever Extraction component integrates with the NVIDIA nv-ingest microservice for data ingestion, processing, and extraction of text files.

The nv-ingest service supports multiple extraction methods for PDF, DOCX, and PPTX file types, and includes pre- and post-processing services like splitting, chunking, and embedding generation. The extractor service's High Resolution mode uses the nemoretriever-parse extraction method for better quality extraction from scanned PDF documents. This feature is only available for PDF files.

The NVIDIA Retriever Extraction component imports the NVIDIA Ingestor client, ingests files with requests to the NVIDIA ingest endpoint, and outputs the processed content as a list of Data objects. Ingestor accepts additional configuration options for data extraction from other text formats. To configure these options, see the component parameters.

tip

NVIDIA Retriever Extraction is also known as NV-Ingest and NeMo Retriever Extraction.

Prerequisites

• An NVIDIA Ingest endpoint. For more information on setting up an NVIDIA Ingest endpoint, see the NVIDIA Ingest quickstart.

• The NVIDIA Retriever Extraction component requires the installation of additional dependencies to your Langflow environment. To install the dependencies in a virtual environment, run the following commands.

  • If you have the Langflow repository cloned and installed from source:
  
  

  _10

  source **YOUR_LANGFLOW_VENV**/bin/activate

  _10

  uv sync --extra nv-ingest

  _10

  uv run langflow run

  
  
  • If you are installing Langflow from the Python Package Index:
  
  

  _10

  source **YOUR_LANGFLOW_VENV**/bin/activate

  _10

  uv pip install --prerelease=allow 'langflow[nv-ingest]'

  _10

  uv run langflow run

  
  

Use the NVIDIA Retriever Extraction component in a flow

The NVIDIA Retriever Extraction component accepts Message inputs, and then outputs Data. The component calls an NVIDIA Ingest microservice's endpoint to ingest a local file and extract the text.

To use the NVIDIA Retriever Extraction component in your flow, follow these steps:

1. Add the NVIDIA Retriever Extraction component to your flow.

2. In the Base URL field, enter the URL of the NVIDIA Ingest endpoint. You can also store the URL as a global variable to reuse it in multiple components and flows.

3. Click Select Files to select a file to ingest.

4. Select which text type to extract from the file: text, charts, tables, images, or infographics.

5. Optional: For PDF files, enable High Resolution Mode for better quality extraction from scanned documents.

6. Select whether to split the text into chunks.

   You can modify the splitting parameters and other hidden settings through the Controls in the component's header menu.

7. Click Run component to ingest the file, and then open the Logs pane to confirm the component ingested the file.

8. To store the processed data in a vector database, add a Vector Store component to your flow, and then connect the NVIDIA Retriever Extraction component's Data output to the Vector Store component's input.

   When you run the flow with a Vector Store component, the processed data is stored in the vector database. You can query your database to retrieve the uploaded data.

   

NVIDIA Retriever Extraction component parameters

The NVIDIA Retriever Extraction component has the following parameters.

For more information, see the NV-Ingest documentation.

Inputs

Name	Display Name	Info
base_url	NVIDIA Ingestion URL	The URL of the NVIDIA Ingestion API.
path	Path	File path to process.
extract_text	Extract Text	Extract text from documents. Default: true.
extract_charts	Extract Charts	Extract text from charts. Default: false.
extract_tables	Extract Tables	Extract text from tables. Default: true.
extract_images	Extract Images	Extract images from document. Default: true.
extract_infographics	Extract Infographics	Extract infographics from document. Default: false.
text_depth	Text Depth	The level at which text is extracted. Options: 'document', 'page', 'block', 'line', 'span'. Default: page.
split_text	Split Text	Split text into smaller chunks. Default: true.
chunk_size	Chunk Size	The number of tokens per chunk. Default: 500. Make sure the chunk size is compatible with your embedding model. For more information, see Tokenization errors due to chunk size.
chunk_overlap	Chunk Overlap	Number of tokens to overlap from previous chunk. Default: 150.
filter_images	Filter Images	Filter images (see advanced options for filtering criteria). Default: false.
min_image_size	Minimum Image Size Filter	Minimum image width/length in pixels. Default: 128.
min_aspect_ratio	Minimum Aspect Ratio Filter	Minimum allowed aspect ratio (width / height). Default: 0.2.
max_aspect_ratio	Maximum Aspect Ratio Filter	Maximum allowed aspect ratio (width / height). Default: 5.0.
dedup_images	Deduplicate Images	Filter duplicated images. Default: true.
caption_images	Caption Images	Generate captions for images using the NVIDIA captioning model. Default: true.
high_resolution	High Resolution (PDF only)	Process PDF in high-resolution mode for better quality extraction from scanned PDF. Default: false.

Outputs

The NVIDIA Retriever Extraction component outputs a list of Data objects where each object contains:

• text: The extracted content.
  • For text documents: The extracted text content.
  • For tables and charts: The extracted table/chart content.
  • For images: The image caption.
  • For infographics: The extracted infographic content.
• file_path: The source file name and path.
• document_type: The type of the document, which can be text, structured, or image.
• description: Additional description of the content.

The output varies based on the document_type:

• Documents with document_type: "text" contain:

  • Raw text content extracted from documents, for example, paragraphs from PDFs or DOCX files.
  • Content stored directly in the text field.
  • Content extracted using the extract_text parameter.

• Documents with document_type: "structured" contain:

  • Text extracted from tables, charts, and infographics and processed to preserve structural information.
  • Content extracted using the extract_tables, extract_charts, and extract_infographics parameters.
  • Content stored in the text field after being processed from the table_content metadata.

• Documents with document_type: "image" contain:

  • Image content extracted from documents.
  • Caption text stored in the text field when caption_images is enabled.
  • Content extracted using the extract_images parameter.