LEANN

LEANN achieves extreme storage efficiency by building a pruned graph during index construction where only high-degree hub nodes retain full embeddings, while low-degree nodes have embeddings discarded. During search, pruned embeddings are recomputed on-demand during graph traversal using the embedding model, trading compute for storage. This approach uses high-degree preserving pruning to maintain search accuracy while eliminating the need to store millions of embedding vectors in full precision.

LEANN provides a backend plugin system that abstracts vector search algorithms, allowing users to swap between HNSW (hierarchical navigable small world graphs for in-memory search), DiskANN (disk-optimized approximate nearest neighbor for large-scale indexing), and IVF (inverted file index for clustering-based search). Each backend implements a common interface for index building, searching, and metadata filtering, enabling performance tuning without changing application code.

LEANN exposes both a Python API (for programmatic use in applications) and a command-line interface (for index building, searching, and management tasks). The API provides high-level abstractions for index creation, document addition, search, and RAG operations, while the CLI enables batch operations and scripting without writing Python code.

LEANN enables building RAG applications over personal data (emails, notes, files, browsing history) with all processing happening locally on the user's device. No data is sent to cloud services unless explicitly configured, and the system provides privacy guarantees through local embedding computation and storage, making it suitable for sensitive personal information.

LEANN can integrate with live data sources (APIs, databases, web services) through MCP tools, allowing RAG queries to incorporate real-time information alongside indexed documents. This enables hybrid RAG that combines static indexed knowledge with dynamic live data, useful for applications requiring current information.

LEANN provides configuration options for tuning index performance across multiple dimensions: backend selection (HNSW, DiskANN, IVF), pruning ratio (controlling storage vs. accuracy tradeoff), distance metrics, and search parameters (ef, num_probes). Users can benchmark different configurations and select optimal settings for their hardware and latency requirements.

LEANN computes embeddings locally using Ollama (for open-source models like Nomic Embed, Llama 2) or via local embedding servers, with optional fallback to OpenAI/Anthropic APIs. The embedding computation layer abstracts provider selection, batching, and caching, allowing users to keep all data on-device while optionally using cloud APIs for specific models. Embeddings are cached after computation to avoid redundant recomputation.

LEANN includes specialized document chunking that parses code using Abstract Syntax Trees (AST) to preserve semantic boundaries (functions, classes, methods) rather than naive line-based or token-based splitting. This enables more accurate semantic search over codebases by ensuring chunks correspond to logical code units, improving retrieval quality for code-specific RAG applications.

LEANN monitors the file system for changes (new files, modified files, deletions) and incrementally updates indices without full rebuilds. The system uses file modification timestamps and content hashing to detect changes, then recomputes embeddings only for modified chunks, reducing index update time from hours to minutes for large document collections.

LEANN supports filtering search results by metadata (document type, date range, tags, custom fields) before or after vector search, and provides configurable distance metrics (cosine similarity, L2 distance, inner product) with optional vector normalization. Metadata filters are applied efficiently during graph traversal to reduce search space, and distance metrics can be swapped per-query without index rebuilds.

LEANN includes a ReAct (Reasoning + Acting) agent implementation that decomposes complex queries into multi-step reasoning chains, using vector search as a tool alongside other capabilities. The agent maintains conversation context, plans actions (search, summarize, retrieve), executes them, and iterates based on results, enabling complex information retrieval tasks beyond simple semantic search.

LEANN exposes a Model Context Protocol (MCP) server that allows Claude and other MCP-compatible clients to query LEANN indices directly from IDEs or Claude Code. This enables developers to use LEANN as a knowledge source within their development environment, retrieving relevant code, documentation, or context without leaving their editor.

LEANN supports task-specific embedding models (e.g., models fine-tuned for code, legal documents, scientific papers) and allows users to define custom prompt templates that modify how text is embedded. This enables optimizing embedding quality for specific domains by using domain-adapted models or prepending task-specific instructions to documents before embedding.

LEANN includes a document loading pipeline that supports multiple formats (PDF, TXT, Markdown, JSON, code files) with format-specific parsing and chunking strategies. The pipeline handles document extraction, text cleaning, and semantic chunking (respecting paragraph/section boundaries), producing chunks optimized for embedding and retrieval.