Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Jev RAG

Local-first knowledge search with BM25 shortlisting and Jev Noul reranking by default, seven selectable pipelines and cited answers.

AppbetaMIT
View on GitHub ↗Live demo ↗

Overview

Jev RAG is an open-source local knowledge search engine. By default SQLite BM25 shortlists passages from a document folder, Jev reranks up to 30 candidates through one batched set of Noul relevance questions with an optional threshold, and the strongest evidence goes with file citations to a configurable answer model that streams a grounded answer. The default path needs no embeddings or vector database; it is local-first rather than offline because selected passages are sent to the Jev and answer-model providers.

Seven pipelines are selectable in one bilingual UI: BM25 plus Jev, agentic lexical multi-query search, embedding hybrid with reciprocal rank fusion, multi-round agentic hybrid, taxonomy-routed hybrid, a unified Jev passage gate and hierarchical Jev line search. Python owns thresholds, ordering, caching and whether evidence reaches the answering model. The project publishes quality, latency, cost and negative results, including a 323-query NFCorpus benchmark, and a BM25-only demo runs without any key.

What the decision model decides

TypeQuestionAnswers
Yes / noIs this passage relevant to the query?yes · no
Yes / noDoes this passage support an answer to the question?yes · no

How it works

  1. Index a local folder with SQLite FTS5 BM25
  2. Shortlist candidates (or use agentic or hybrid retrieval)
  3. Rerank with batched Jev Noul questions
  4. Pass top evidence with citations to the answer model
  5. Stream a grounded answer

Strengths

  • No vector database required by default
  • Published benchmarks including negative results
  • Keyless BM25 demo
  • Several pipelines to compare

Caveats

  • Selected passages leave the machine
  • Needs an OpenRouter key for full pipeline
  • Unofficial project

Getting started

pip install -e '.[documents]', add OPENROUTER_API_KEY to .env, then jev-rag serve and open http://127.0.0.1:8765

At a glance

Type
App
Status
beta
Language
Python
Runtime
Local web app and CLI
Has code
yes
Has tests
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • Retrieval and RAG: shortlists local document passages with BM25, reranks up to 30 candidates through one batched set of Jev `Noul` relevance questions with an optional threshold, and passes the strongest evidence with source citations to a configurable answer model, backed by a runnable local demo and a public 323-query NFCorpus benchmark.

    yibie/awesome-jev

  • Unofficial local-first knowledge search app using SQLite BM25 or agent-planned lexical retrieval, Jev `Noul` judgments for evidence reranking, and a published reproducible NFCorpus evaluation; no vector database is required by default

    AnotiaWang/awesome-decision-models

  • Local knowledge search with three modes: BM25 + Jev by default, agentic multi-query lexical search + Jev without vectors, and optional BM25 + embeddings/RRF + Jev, with grounded streaming answers.

    logicrw/awesome-jev-projects

  • Jev RAG (Jev-RAG / JevRAG): open-source local knowledge search with BM25 + Jev…

    wh000wh000/awesome-jev-live

  • Local knowledge search where SQLite BM25 or agent-planned lexical retrieval shortlists passages, Jev answers batched `Noul` questions about relevance and answer support, and Python owns thresholds, ordering, caching, and whether evidence reaches the answering model.

    cobanov/awesome-jev

  • Jev RAG (Jev-RAG / JevRAG): open-source local knowledge search with BM25 + Jev reranking, agentic and hybrid retrieval, cited answers, and reproducible benchmarks.

    hellogumbo/awesome-jev

  • Local knowledge search with seven selectable pipelines, BM25 plus Jev reranking by default, streaming cited answers over your own document folder.

    MrJev/awesome-jev

Frequently asked questions

What does Jev decide?

Whether each shortlisted passage is relevant and supports an answer, used to rerank evidence.

Do I need a vector database?

No. The default pipeline is BM25 plus Jev.

Is it fully offline?

No. Selected passages are sent to Jev and the answer model.

Related projects

  • jev-rerank★ 17

    Jev as a calibrated reranker: query plus up to 30 documents in one call, one relevance Noul per document, benchmarked on BEIR.

    ProjectbetaPython
  • Vector Graph RAG

    Milvus-based multi-hop graph RAG library where two-stage Jev Noul scoring selects relations and reranks passages.

    LibrarybetaPython
  • Jev-Reranking

    Compares Jev's zero-shot reranking with specialist rerankers on MS MARCO passage and document benchmarks and a TREC WSJ transfer test.

    ProjectresearchPython
  • JevRanker

    Open Jev-style 0.6B decision model plugged into BlitzRank and Reranker-Guided Search, scoring k candidates per match in one pass.

    ServiceresearchPython
  • jevector

    Profile matching that encodes preference criteria as Clef or Jev decision vectors, retrieves with HNSW and checks rules both ways.

    BenchmarkresearchPython
  • Jev vs embeddings for page selection

    Sayora write-up comparing Jev with production embeddings for picking the right website page to answer caller questions.

    Articleproduction