A self-hosted RAG service. You upload documents, ask questions, and get streamed answers that cite the passages they came from. If the documents don’t support an answer, it says so instead of guessing.
- Hybrid retrieval with dense vectors and BM25 fused by reciprocal rank. It abstains when the evidence is weak and audits every citation it emits.
- A golden set of 41 questions with labelled evidence, including multi-hop, cross-language and adversarial cases, scored with Recall@k, nDCG and MRR and bootstrap confidence intervals. Embedding models are chosen by A/B on that set.
- About 300 tests behind an 85% branch-coverage gate, Playwright end-to-end runs, a backup and restore round trip, and multi-arch images with SBOM and provenance.
- Python
- FastAPI
- LangChain
- Qdrant
- Ollama