NeMo Retriever is NVIDIA’s open-source library for extracting usable content from documents at scale — the ingestion layer that feeds retrieval-augmented generation (RAG) systems. It pulls text, tables, charts and images out of PDFs and other documents, along with their metadata, and prepares them for embedding and search in downstream generative AI applications. The GitHub project has around 3,000 stars and is actively maintained by NVIDIA.
Enterprise RAG systems are only as good as what goes into them, and most enterprise knowledge is trapped in messy documents. NeMo Retriever is built as a scalable, performance-oriented extraction pipeline: documents are ingested through a library API for smaller jobs, or as a microservice architecture for production volumes. Extraction is powered by NVIDIA NIM microservices — specialized models for page element detection, table and chart understanding, and OCR — while development setups can also use local Hugging Face models. The result is structured, contextualized content that embedding models and vector databases can actually use, instead of the garbled text that naive PDF parsing produces.
NeMo Retriever is free and open source under the Apache-2.0 license. For production use it can call NVIDIA NIM microservices, which are licensed separately by NVIDIA, but the library itself costs nothing.
NeMo Retriever is for AI engineers building enterprise RAG pipelines who need high-fidelity document extraction — especially teams already invested in NVIDIA GPUs and the NeMo/NIM stack.