Jev RAG

No. The default pipeline is vector-free: SQLite FTS5/BM25 retrieves candidates, Jev reranks them, and an answer model streams cited responses. Optional hybrid modes cache a normalized embedding matrix locally without requiring a separate vector database.

The full BEIR NFCorpus test split: 3,633 documents and 323 queries, with the same corpus, queries, qrels and metric implementation across rows. Latency covers retrieval and reranking only, not one-time indexing or final answer generation.

BM25 plus Jev is the default, alongside agentic lexical, embedding hybrid, multi-round Agentic Hybrid, taxonomy-routed hybrid, a unified Jev Passage Gate and hierarchical Jev Line Search, all selectable in one UI. The best measured row is Agentic Hybrid with Jev and retrieval rank fusion at 0.450750 nDCG@10, using a dev-selected local RRF at Jev weight 1.0 and retrieval weight 0.25 that adds no provider call.

The unified Passage Gate is a disclosed negative: at cookbook-style fixed thresholds it excluded 13,631 of 16,150 candidates, 84.4%, and scored below both bare Hybrid at 0.396712 and Hybrid plus Jev at 0.444327, so it is kept for injection screening rather than quality. Line Search had the best first hit at nDCG@1 of 0.554180 but a full-corpus cold provider cost of $4.55.

Local-first but not offline: selected passages go to the configured Jev and answer-model providers, and hybrid mode also sends document text and queries to an OpenRouter embedding model. The optional embedding matrix stays local. For a key-free trial, local_kb.py indexes ./docs into a SQLite FTS5 database and search --no-jev returns ranked passages with no install, key or model request.

Python 3.9+ is the floor, then a venv, pip install -e '.[documents]', an OPENROUTER_API_KEY in .env and jev-rag serve, which opens http://127.0.0.1:8765 and discovers ~/Documents by default. It indexes Markdown, text, HTML, JSON, CSV, YAML, DOCX and PDF with CJK bigram tokenization. It is MIT licensed but marked alpha, so APIs and storage schemas may change before 1.0.

Not as a rank. On the MTEB NFCorpus page observed on 2026-09-26, inserting 0.450750 would place the run at roughly 4th of 251 results, top 1.6%, but the README calls that an unofficial comparison: the multi-stage pipeline has not been submitted to MTEB and its top-50 configuration was evaluated on the same test set. Provider latency also moves substantially over time.