hRAG

hRAG provides a self-hosted hybrid RAG solution that integrates BM25 lexical search and vector embeddings within a single Postgres database. This architecture eliminates the need for separate search clusters, simplifying operations and reducing infrastructure costs. The platform includes a cross-encoder reranker to improve retrieval accuracy, which is offered as an optional feature due to its latency impact.

The system emphasizes transparency and verifiability, with every claim backed by public benchmark scores and reproducible results. It features grounded answers, refusing to hallucinate when information is not found, and provides streaming citations that link directly to source chunks. Tenant isolation is enforced using Postgres row-level security, ensuring data privacy for multiple users or sandboxes.

hRAG is designed for cost-efficiency, running on a five-node Kubernetes cluster on Hetzner for approximately €116 per month. It utilizes small, budget-friendly models for embedding and answering, benchmarked against more expensive cloud alternatives. The entire codebase is MIT licensed, and deployment articles are provided to guide users through self-hosting.

The platform is officially scored on EnterpriseRAG-Bench, achieving a #9 ranking with strong performance in correctness, completeness, and document recall, outperforming Vertex AI Search and NVIDIA in some metrics. Users can try a 512K-document playground without login, or sign in with Google/GitHub for a private sandbox with limited document capacity and token budget.