What is NVIDIA-Ingest
NVIDIA-Ingest is a collection of microservices open-sourced by NVIDIA for parsing complex, messy unstructured PDFs and other enterprise documents. NVIDIA-Ingest can convert documents into metadata and text, making it easy to embed into retrieval systems. NVIDIA-Ingest supports multiple document formats such as PDF, Word, PowerPoint, and images, and provides multiple extraction methods to facilitate trade-offs between throughput and accuracy. NVIDIA-Ingest supports pre-processing and post-processing operations such as text splitting, conversion, filtering, embedding generation, and image storage. NVIDIA-Ingest is based on parallelized document processing to improve extraction efficiency, supports embedding extracted content into vector databases such as Milvus, and is suitable for large-scale document processing and generative applications.
Main features of NVIDIA-Ingest
- Multi-format document support: Supports parsing multiple complex enterprise document formats such as PDF, Word (Docx), PowerPoint (Pptx), and images.
- Multi-method extraction: Supports multiple extraction methods to facilitate trade-offs between throughput and accuracy. For example, PDF documents support extraction with pdfium, Unstructured.io, and Adobe Content Extraction Services.
- Content classification and extraction: Classifies document content into text, tables, charts, and images, and extracts these contents separately. Optical Character Recognition (OCR) technology is used to further contextualize the extracted content and convert it into a well-defined JSON schema.
- Parallel processing: Supports splitting documents into pages and processing content extraction for each page in parallel, improving processing efficiency.
- Pre-processing and post-processing: Supports multiple pre-processing and post-processing operations, including text splitting and chunking, content conversion, filtering, embedding generation, and image storage.
Technical principles of NVIDIA-Ingest
- Microservice architecture: Based on a microservice architecture, each microservice is responsible for specific processing tasks such as text extraction, image extraction, table extraction, etc., improving the system's scalability and flexibility.
- GPU acceleration: Based on NVIDIA's GPU technology, especially H100 and A100 GPUs, it accelerates document parsing and content extraction. The parallel computing capability of GPUs significantly improves processing efficiency, especially when processing large numbers of documents.
- Optical Character Recognition (OCR): Uses OCR technology to convert image and table content in documents into readable text. NVIDIA-Ingest integrates multiple OCR engines, such as PaddleOCR, to improve the accuracy and efficiency of text recognition.
Project address of NVIDIA-Ingest
- GitHub repository：https://github.com/NVIDIA/nv-ingest
Application scenarios of NVIDIA-Ingest
- Enterprise content management: Converts paper documents, PDFs, Word, and PowerPoint, etc. into searchable, editable digital formats, supporting knowledge sharing and collaboration.
- Intelligent customer service systems: Parses documents uploaded by users, extracts key information, and generates automatic responses, improving customer service efficiency and user satisfaction.
- Legal and compliance field: Parses contracts and legal documents, extracts key terms and conditions, and supports compliance checks, risk assessment, and case management.
- Financial industry: Parses financial reports, contracts, and market research documents, extracts key data, and supports risk assessment, compliance monitoring, and customer due diligence.
- Healthcare: Convert medical record documents into structured data, supporting electronic medical record management, clinical research, and medical imaging analysis.