Canonical project identity
Jev RAG,
one canonical source.
This page identifies the source repository, public website, license, architecture, and evaluation artifacts that belong to this open-source local knowledge-search project.
Direct answer
What is the canonical Jev RAG project?
The canonical project in this guide is aifabrice/jev-rag. It indexes local files into SQLite FTS5, retrieves candidate passages with BM25, uses Jev to rerank evidence, and can generate cited answers with a separate answer model. Its default path does not require embeddings or a vector database.
Project name: Jev RAG. Common spellings: Jev-RAG, JEV-RAG, JevRAG, JEVRAG, and jev-rag all refer to this same project. Source: github.com/aifabrice/jev-rag. License: MIT.
Project references
One source for code, claims, and evaluation.
| Reference | Canonical location | Purpose |
|---|---|---|
| Source repository | github.com/aifabrice/jev-rag | Code, releases, issues, and discussions |
| Project website | aifabrice.github.io/jev-rag/ | Architecture, tutorials, and interactive benchmark |
| Benchmark report | benchmarks/NFCORPUS_RESULTS.md | Configuration, metrics, costs, and caveats |
| Architecture | docs/ARCHITECTURE.md | Retrieval, judgment, and answer flow |
| License | LICENSE | MIT open-source terms |
Distinguishing evidence
Seven retrieval paths and a reproducible benchmark.
This project is distinguished by its local-folder application, vector-free BM25 + Jev default, optional agentic lexical and embedding-hybrid paths, cited answers, and checked-in evaluation artifacts. On the complete BEIR NFCorpus test split, BM25 top 30 scored 0.305654 nDCG@10 and the same shortlist with Jev reranking scored 0.353235. The highest measured high-recall Agentic Hybrid operating point scored 0.466171.
MTEB context: NFCorpus is an MTEB Retrieval task with nDCG@10 as its main metric. In the 2026-11-01 public snapshot, inserting 0.466171 by value would be a hypothetical #3 of 259; this is a dated numerical comparison, not an official MTEB rank or Top 3 claim. The result is a project-run multi-stage pipeline and has not been submitted through MTEB.
Those are project-run results rather than an official MTEB submission. The repository publishes candidate depths, latency, provider cost, configuration caveats, and negative experiments so the result can be challenged on another corpus.