NVIDIA RAG Blueprint

NVIDIA RAG Blueprint is a reference framework for developers building AI question answering over enterprise data. It retrieves relevant material from your documents to ground answers and can cite the retrieved sources.

• Document extraction: NeMo Retriever extracts text, tables, charts, infographics and audio, with support for custom metadata.
• Search across collections: Combine semantic and keyword search, filter results by metadata and rerank retrieved passages before generating an answer.
• Agentic questions: A LangGraph pipeline breaks complex or ambiguous questions into parallel searches, combines the findings and can verify the result across documents.
• Answers and summaries: Use optional vision language models for multimodal answers, summarize selected document pages and apply programmable content safety guardrails.
• Chat and evaluation: A reference UI supports multi-turn conversations and separate sessions, while RAGAS evaluation scripts help assess answer quality.

The framework runs through local Docker deployments or on Kubernetes, including Red Hat OpenShift. It uses NVIDIA NIM services for generation, embeddings and reranking, including Nemotron models, and supports GPU-accelerated indexing and retrieval. Elasticsearch is the default vector database; Milvus is an alternative. OpenAI-compatible APIs and a Python library mode support integration into other applications.

You can self-host the model services or connect to NVIDIA-hosted endpoints. With hosted endpoints, the affected model requests go to NVIDIA's service. The blueprint code is open source under Apache 2.0; its models have separate NVIDIA and Llama license terms.