LEANN-RS

"Rewrite it in Rust" version of LEANN - Lightweight Embedding-based Approximate Nearest Neighbor search.

A single-binary CLI for building, searching, and querying vector indexes for RAG applications. No Python runtime required.

Features

• Single binary - No Python runtime, no dependencies
• Fast - Native Rust with optimized HNSW/DiskANN search
• Portable - Cross-platform (Linux, macOS, Windows)
• Compatible - Reads Python LEANN index format
• Hybrid search - Combine vector + BM25 keyword search
• Metadata filtering - Filter results by document attributes
• ReAct agent - Multi-turn reasoning with tool use
• HTTP server - REST API for integration

Installation

# Build from source
cargo build --release

# Binary at target/release/leann (~4MB)

Quick Start

# Build an index from documents
leann build my-docs --docs ./documents --embedding-mode ollama --embedding-model nomic-embed-text

# Search
leann search my-docs "How does authentication work?"

# Ask questions (RAG)
leann ask my-docs "Explain the architecture" --llm ollama --model qwen3:8b

# Interactive mode
leann ask my-docs --interactive

Commands

Build Index

# Basic build with OpenAI embeddings
leann build my-docs --docs ./documents

# Use Ollama for local embeddings
leann build my-docs --docs ./documents --embedding-mode ollama --embedding-model nomic-embed-text

# Custom options
leann build my-code --docs ./src \
  --file-types ".rs,.py,.ts" \
  --doc-chunk-size 512 \
  --graph-degree 48

Search

# Basic search
leann search my-docs "vector database"

# With metadata filtering
leann search my-docs "authentication" -f "source:*.rs"

# Hybrid search (vector + BM25)
leann search my-docs "user login" --hybrid

# JSON output
leann search my-docs "query" --format json

Ask (RAG)

# Single question
leann ask my-docs "How does caching work?"

# Use different LLM providers
leann ask my-docs "question" --llm openai --model gpt-4o
leann ask my-docs "question" --llm anthropic --model claude-3-5-sonnet-20241022
leann ask my-docs "question" --llm ollama --model qwen3:8b

# Interactive chat
leann ask my-docs --interactive

ReAct Agent

Multi-turn reasoning with search tool:

leann react my-docs "What are all the ways errors are handled?"

# Show reasoning trace
leann react my-docs "Compare feature X and Y" --verbose --max-steps 10

HTTP Server

# Start server
leann serve my-docs --port 8080 --cors

# API endpoints:
# POST /search - Search the index
# GET  /info   - Index information
# GET  /health - Health check

Manage Indexes

# List all indexes
leann list

# Detailed info
leann list --detailed

# Remove an index
leann remove my-docs

Optional Features

# Build with PDF support
cargo build --release --features pdf

# Build with HTTP server
cargo build --release --features server

# Build with DiskANN backend
cargo build --release --features diskann-backend

# Build with all features
cargo build --release --features full

Search Options

Metadata Filtering

Filter by document metadata:

# By file extension
leann search my-docs "query" -f "source:*.rs"

# By type
leann search my-docs "query" -f "type=code"

# Numeric comparison
leann search my-docs "query" -f "lines>100"

Supported operators: =, !=, >, >=, <, <=, : (glob patterns)

Hybrid Search

Combine vector similarity with BM25 keyword matching:

leann search my-docs "exact function name" --hybrid --hybrid-alpha 0.5

Alpha controls the balance: 1.0 = pure vector, 0.0 = pure BM25.

Index Compatibility

LEANN-RS reads Python LEANN indexes:

• .passages.jsonl - Text passages
• .passages.idx.json - JSON offset map
• .index - Vector index (usearch format)
• .meta.json - Metadata
• .ids.txt - ID mapping

Environment Variables

Binary Sizes

Architecture

src/
├── cli/           # Commands (build, search, ask, react, serve)
├── index/         # Index management, BM25, filtering
├── backend/       # HNSW (usearch), DiskANN
├── embedding/     # OpenAI, Ollama providers
└── llm/           # OpenAI, Ollama, Anthropic

License

MIT