Open source · MIT · Python
Jev RAG,
without the vector tax.
Jev RAG searches local documents through seven selectable paths: BM25 + Jev, agentic lexical search, embedding hybrid retrieval, multi-round Agentic Hybrid, corpus-taxonomy routing, a unified Jev Passage Gate, or hierarchical Jev Line Search. Embeddings remain optional.
$
git clone https://github.com/aifabrice/jev-rag.gitPublic benchmark
See what each stage adds.
Complete BEIR NFCorpus test split. Select a pipeline to compare its recorded retrieval quality against the BM25 baseline.
Agentic Hybrid high-recall top 200 + Jev0.466171nDCG@10
Median latency
p95 latency
10-query unified cold audit; caches disabled; Jev batch size 100
Benchmark-only high-recall operating point: 200 candidates from both retrieval branches, then Jev batches of 100.
Measured, not marketed. These are project-run results, not an official MTEB submission. The relevant MTEB NFCorpus Retrieval task uses nDCG@10; in the 2026-11-01 public snapshot, inserting this run's 0.466171 numerically would be a hypothetical #3 of 259, not an official MTEB Top 3. Latency covers retrieval and reranking, excludes one-time corpus indexing and final answer generation, and remains provider-dependent. Ordinary shortlist rows use one unified 10-query cold audit with caches disabled; historical specialized rows are labeled and are not directly comparable. Near-duplicate and negative experiments remain in the full report for reproducibility. Read the seven-pipeline field report → Inspect the complete benchmark artifacts → Inspect latency provenance →
Architecture
Start simple. Add semantics only when they earn their keep.
LocalYour documentsPDF · DOCX · Markdown · HTML
Index
DefaultSQLite FTS5 / BM25Fast, inspectable lexical retrieval
OptionalAgentic, embedding, or bothTwo-round search planning and semantic recall can be fused
ExperimentalCorpus taxonomy routingHybrid top 50 plus routed leaf-node candidates
ExperimentalUnified Jev Passage GateRelevance, evidence, conflict and injection routing
ExperimentalTwo-level Jev Line SearchParallel 255-passage windows, then global Choice
Rerank
EvidenceJevScore and order candidate passages
Answer
GroundedMiniMaxStreaming response with citations
Deliberate trade-offs
No architecture wins every query.
Exact terms, IDs and names
BM25 is fast, local, transparent, and excellent when question language overlaps the source.
Synonyms and paraphrases
The default path can miss semantic matches. Agentic search or hybrid retrieval can widen recall.
Know what leaves the machine
Default retrieval stays local before reranking. Line Search is different: a bounded representation of every indexed passage is sent to Jev in windows, making it the most expensive mode.
Direct answer
What is Jev RAG?
Jev RAG is an open-source, local-first knowledge search application. It indexes files into SQLite FTS5, retrieves passages with BM25, uses the Jev model to rerank evidence, and asks MiniMax to produce a cited answer. Its default path does not require embeddings or a vector database.
Use BM25 for exact terms, Agentic for query reformulation, and hybrid retrieval for semantic recall. Use experimental Line Search when you want to measure Jev-only full-corpus retrieval and accept its larger cost and remote-data boundary.
Try it on your own corpus