Crate leann_core 

LEANN — Lightweight Embedding-Augmented Nearest Neighbors.

Build, search, and manage vector indexes with optional graph-based embedding recomputation for 97% storage reduction.

Quick start

use leann_core::{LeannBuilder, LeannSearcher};

// Build an index
let mut builder = LeannBuilder::new("model-name", Some(384), "sentence-transformers");
builder.add_text("Hello world", Default::default());
builder.build_index(&path, &provider)?;

// Search
let searcher = LeannSearcher::open(&path)?;
let results = searcher.search("hello", 5)?;

Re-exports

pub use backend::BackendConfig;

pub use backend::BackendIndex;

pub use builder::LeannBuilder;

pub use chat::LeannChat;

pub use index::IndexMeta;

pub use passages::Passage;

pub use passages::PassageManager;

pub use search_result::SearchResult;

pub use searcher::LeannSearcher;

pub use searcher::SearcherOptions;

Modules

backend

Backend abstraction for LEANN index engines.

builder

chat

chunking

document_loaders

Document loaders for various file formats.

embedding

hnsw

index

multi_vector

Multi-vector (ColBERT) index support.

passages

react_agent

search_result

searcher

sources_manifest

Sources manifest for incremental build support.

sync

Structs

BM25Scorer

BM25 scoring for keyword-based search.

MetadataFilterEngine

Engine for evaluating metadata filters against search results.

Type Aliases

FilterSpec

A filter specification: maps operator names to expected values. Example: {"<=": 5} or {"in": ["fiction", "drama"]}

MetadataFilters

Metadata filters: maps field names to their filter specifications. Example: {"chapter": {"<=": 5}, "tags": {"in": ["fiction", "drama"]}}