How aifabrice/jev-rag works
Jev RAG is a local-first knowledge search engine: users index a document folder, select a retrieval pipeline, and receive ranked passages or a streamed answer with file citations. The runtime is primarily implemented in local_kb.py, with taxonomy.py providing taxonomy construction and routing. The README documents the seven pipeline capabilities and their broad stages; source excerpts are partial, so internal calls beyond the shown taxonomy reference are not asserted.
Main entry points: local_kb.py Jev reranking · taxonomy.py Taxonomy routing
Components 18
User entry points
- CLI
local_kb.py - Web interface
local_kb.py
Local knowledge
- Document folder
- Document ingestion
local_kb.py - BM25 indexing
local_kb.py - SQLite FTS5
local_kb.py
Retrieval pipelines
- BM25 retrieval
local_kb.py - Agentic lexical
local_kb.py - Embedding hybrid
local_kb.py - Agentic hybrid
local_kb.py - Taxonomy routing
taxonomy.py - Passage Gate
local_kb.py - Line Search
local_kb.py
Ranking and answers
- Jev reranking
local_kb.py - Answer generation
local_kb.py - Ranked passages and cited answer
Other components
- User
- Configured model providers
Connections 29
- User → CLI: runs commands
README.md - User → Web interface: uses interface
README.md - Document folder → Document ingestion: reads files
README.md - Document ingestion → BM25 indexing: indexes content
README.md - BM25 indexing → SQLite FTS5: stores passages
README.md - SQLite FTS5 → BM25 retrieval: searches passages
README.md - SQLite FTS5 → Agentic lexical: searches lexical terms
README.md - SQLite FTS5 → Embedding hybrid: provides BM25 candidates
README.md - SQLite FTS5 → Agentic hybrid: provides BM25 candidates
README.md - Embedding hybrid → Taxonomy routing: routes candidates
README.md - SQLite FTS5 → Passage Gate: provides BM25 candidates
README.md - Embedding hybrid → Passage Gate: provides hybrid candidates
README.md - Configured model providers → Agentic lexical: plans searches
README.md - Configured model providers → Embedding hybrid: creates embeddings
README.md - Configured model providers → Agentic hybrid: plans and embeds
README.md - Configured model providers → Taxonomy routing: embeds corpus
README.md - Configured model providers → Line Search: scores windows
README.md - BM25 retrieval → Jev reranking: reranks candidates
README.md - Agentic lexical → Jev reranking: reranks candidates
README.md - Embedding hybrid → Jev reranking: reranks candidates
README.md - Agentic hybrid → Jev reranking: reranks candidates
README.md - Taxonomy routing → Jev reranking: reranks candidates
README.md - Passage Gate → Answer generation: screens evidence
README.md - Line Search → Answer generation: selects passages
README.md - Jev reranking → Answer generation: supplies ranked evidence
README.md - Configured model providers → Jev reranking: scores candidates
README.md - Configured model providers → Answer generation: generates answer
README.md - Answer generation → Ranked passages and cited answer: streams cited answer
README.md - Jev reranking → Ranked passages and cited answer: returns ranked passages
README.md