Introduction

bigRAG is an open-source RAG platform with Turbopuffer at the retrieval layer. It provides a complete pipeline for document ingestion, chunking, embedding, and search behind a REST API, TypeScript/Python SDKs, an admin UI, and an MCP server.

Key Features

• End-to-end RAG pipeline — upload documents, auto-chunk, embed, search, and rerank in one platform
• Any document format — PDF, DOCX, PPTX, XLSX, HTML, Markdown, CSV/TSV/XML/JSON, and scanned images (OCR) via Docling
• Any embedding model — OpenAI, Cohere, Voyage, or openai_compatible gateways (Ollama, vLLM, TEI, LiteLLM, Azure, Bedrock) configured per collection
• Turbopuffer search backend — vectors, chunk text, payload metadata, BM25 keyword search, and hybrid retrieval in one managed backend
• Namespace isolation — every collection maps to a Turbopuffer namespace for scoped writes, exports, deletion, and retrieval
• Three search modes — semantic, keyword, and hybrid search with optional Cohere reranking
• First-class auth — admin accounts, session cookies, scoped bigrag_sk_… API keys, and full audit/access logs
• Operational features — batch ingestion, S3/R2 prefix sync, webhooks with HMAC signatures, metadata-schema validation, an evaluation runner, and per-collection query analytics
• Admin UI — first-run setup, collection browser, stateless chat, connector management, access logs, MCP credentials, and API-key minting
• Self-hostable control plane — Docker Compose for the API, admin UI, Postgres, and Redis
• MIT licensed — run it anywhere, forever free

Architecture

Explore