How nvidia-ai-blueprints/rag works
NVIDIA RAG Blueprint combines a reference chat and collection-management UI with API services for document ingestion and grounded question answering. The principal workflow is document upload and multimodal extraction into a pluggable vector store, followed by query processing, retrieval/reranking, and answer generation. Agentic planning, multimodal answering, reflection, summaries, safety controls, and observability extend that flow. The README documents these capabilities; source excerpts are partial, so links between unsampled internals are limited to documented responsibilities.
Main entry points: rag_server/main.py RAG API · ingestor_server/main.py Ingestion API · App.tsx Reference UI · vdb_base.py Vector store
Components 25
User Experience
- Reference UI
frontend/src/App.tsx - Chat workflow
frontend/src/pages/Chat.tsx - Collection management
frontend/src/pages/NewCollection.tsx - Settings
frontend/src/pages/SettingsPage.tsx
Ingestion
- Ingestion API
src/nvidia_rag/ingestor_server/main.py - Ingestion task state
src/nvidia_rag/ingestor_server/ingestion_state_manager.py - Document extraction
src/nvidia_rag/ingestor_server/nvingest.py - Ingest schema adapter
src/nvidia_rag/ingestor_server/nemo_retriever/ingest_schema_manager.py - Document summaries
src/nvidia_rag/utils/summarization.py
Query and Retrieval
- RAG API
src/nvidia_rag/rag_server/main.py - Query processing
src/nvidia_rag/rag_server/query_decomposition.py - Agentic RAG
src/nvidia_rag/rag_server/agentic_rag/agentic_rag.py - Retrieval and reranking
src/nvidia_rag/utils/reranker.py
Answering
- Response generation
src/nvidia_rag/rag_server/response_generator.py - Multimodal answering
src/nvidia_rag/rag_server/vlm.py - Reflection
src/nvidia_rag/rag_server/reflection.py - Optional guardrails
Storage and Operations
- Vector store
src/nvidia_rag/utils/vdb/vdb_base.py - Elasticsearch adapter
src/nvidia_rag/utils/vdb/elasticsearch/elastic_vdb.py - Milvus adapter
src/nvidia_rag/utils/vdb/milvus/milvus_vdb.py - NVIDIA NIM services
- Telemetry and tracing
src/nvidia_rag/utils/observability/tracing/instrumentation.py
Other components
- User
- Enterprise documents
- API client
Connections 30
- User → Reference UI: uses
- Reference UI → Chat workflow: routes to
- Reference UI → Collection management: routes to
- Reference UI → Settings: routes to
- Chat workflow → RAG API: submits query
- API client → RAG API: calls API
- RAG API → Query processing: processes query
- Query processing → Retrieval and reranking: prepares search
- RAG API → Agentic RAG: dispatches agentic
- Agentic RAG → Retrieval and reranking: executes searches
- Retrieval and reranking → Vector store: searches
- Elasticsearch adapter → Vector store: implements
- Milvus adapter → Vector store: implements
- RAG API → Response generation: generates answer
- Response generation → NVIDIA NIM services: requests inference
- RAG API → Multimodal answering: dispatches multimodal
- Multimodal answering → NVIDIA NIM services: requests vision inference
- RAG API → Reflection: invokes reflection
- RAG API → Optional guardrails: filters content
- Enterprise documents → Ingestion API: uploads
- User → Collection management: configures collection
- Collection management → Ingestion API: creates collection
- Ingestion API → Ingestion task state: tracks progress
- Ingestion API → Document extraction: dispatches extraction
- Document extraction → NVIDIA NIM services: extracts content
- Document extraction → Ingest schema adapter: adapts records
- Ingest schema adapter → Vector store: writes records
- Ingestion API → Document summaries: requests summaries
- RAG API → Telemetry and tracing: emits traces
- Ingestion API → Telemetry and tracing: emits traces