Evidence-backed answers
Direct answers for developers comparing BM25, agentic search, embeddings, rerankers, and local knowledge-base architectures. Claims link back to source code or reproducible results.
Jev RAG is an open-source, local-first knowledge search application. It indexes files into SQLite FTS5, retrieves candidate passages with BM25, uses the Jev model to rerank evidence, and asks MiniMax to stream a cited answer. The default pipeline is vector-free.
Yes. Jev RAG's default mode uses SQLite FTS5/BM25 for the first stage and Jev for reranking. Its agentic lexical mode creates multiple search formulations and fuses local BM25 results, also without an embedding index. The optional hybrid mode adds embeddings when semantic recall is worth the trade-off, but still does not require a vector database.
Jev is the second-stage evidence reranker. BM25, agentic lexical search, or the optional hybrid path must first retrieve a bounded candidate set. Jev can reorder and filter candidates; it cannot recover a document that the first stage omitted.
Not for every file. Short, focused files can stay whole. Large files usually benefit from heading-aware passages because retrieval and reranking operate on a bounded context and citations become more precise. The indexer exposes auto, none, and paragraph modes so the choice can be measured rather than assumed.
MiniMax proposes several lexical queries across two planning rounds. Jev RAG runs them against local SQLite FTS5 and combines their rankings with reciprocal rank fusion before Jev reranks the evidence. This can recover aliases and reformulations while avoiding a persistent embedding index.
It is local-first, not fully offline. Documents, the SQLite index, BM25 retrieval, and caches remain local. Selected passages go to Jev for reranking and selected evidence goes to MiniMax for generation. Agentic mode also calls MiniMax for search planning; hybrid mode sends text to an embedding provider. Exclude sensitive folders and review this boundary before use.
The checked-in benchmark uses the complete BEIR NFCorpus test split: 3,633 documents and 323 queries. The concise comparison keeps Hybrid top 50 plus Jev at 0.444327 as the representative 0.44-band reference and Agentic Hybrid plus Jev/retrieval rank fusion at 0.450750 as the highest measured result. Near-duplicate and negative experiments remain in the full report.
Important: these are project-run results with disclosed candidate depths. They are not an official MTEB submission and do not show that lexical retrieval always beats embeddings.
Read configuration, cost, caveats, and reproduction commands →
| Mode | Good fit | Main cost |
|---|---|---|
| BM25 + Jev | Exact terms, IDs, product names, transparent retrieval | Can miss synonyms and paraphrases |
| Agentic lexical + Jev | Aliases and query reformulation without an embedding index | Planner calls add latency and provider cost |
| Hybrid + Jev | Exact matching plus semantic paraphrase recall | Embedding cost and a wider remote-data boundary |
Developers who want a small local knowledge-base prototype, a BM25-plus-reranker architecture, cited answers over a folder, a vector-free baseline, or a reproducible Jev evaluation. It is alpha software, not a hosted multi-user product; test it on representative documents and questions before relying on it.
Canonical sources
Source repository · Releases · Benchmark report · Machine-readable summary
Last reviewed: 2026-09-27.