hRAG
Hybrid RAG on €116/month of Hetzner, officially benchmarked
Last checked:
What is hRAG?
hRAG is a retrieval-augmented generation platform that answers questions by searching document collections using hybrid search (BM25 + vectors) and cross-encoder reranking. It publishes transparent benchmark scores, open-source code, and costs €116/month, targeting organizations needing reproducible, cost-effective RAG without vendor lock-in.
hRAG pricing
Pricing model: Freemium
€116/month for complete five-node cluster (ingest, embed, rerank, answer services). Public playground free without login. Private sandbox free with Google/GitHub sign-in (capped at 10 documents, 20 pages each).
hRAG pros
- Hybrid search combining BM25 and vector retrieval with cross-encoder reranking for higher-quality answers
- Transparent and reproducible: ranked #9 on EnterpriseRAG-Bench with published results and MIT-licensed open-source code
- Extremely cost-effective at €116/month compared to cloud alternatives costing 100× more
- Grounded answering refuses hallucinations: 100% accurate on info-not-found questions where competitors hallucinate
hRAG cons
- Cross-encoder reranking adds 16-second latency, so it ships as an optional checkbox rather than default behavior
- Private sandbox limited to 10 documents, 20 pages each, with daily token budget (testing-only scale)
- Requires self-hosting or Kubernetes management; no fully-managed SaaS option available
Frequently asked questions about hRAG
What documents can I search in hRAG?
Public playground has 512,000 company documents (Slack, email, wiki, tickets) from EnterpriseRAG-Bench. Private sandboxes let you upload your own documents (up to 10 docs, 20 pages max, with daily token limits).
How does hRAG rank compared to other RAG solutions?
Ranked #9 on EnterpriseRAG-Bench (44.74 overall) with document recall of 69.65, outperforming Vertex AI (61.76) and Azure (64.25). All results are published and independently reproducible.
Is hRAG open source and self-hostable?
Yes, MIT licensed. All code, deployment guides, and benchmark reproduction steps are published so you can self-host and verify performance claims yourself.
What's included in the €116/month pricing?
A complete five-node Kubernetes cluster on Hetzner with PostgreSQL, BM25 search, vector embeddings, row-level security, and four core services. Individual answers cost fractions of a cent.