Neura Market
LEANN logo

LEANN

Free

[MLsys2026]: RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.

Search ToolsFreeFree tier
Type
Open Source
Visit Website GitHub (11.0k stars)

About LEANN

LEANN is a storage-efficient vector index designed to address the high storage overhead of traditional vector indices used in embedding-based search applications like retrieval-augmented generation (RAG) and recommendation. Instead of storing high-dimensional embeddings, LEANN recomputes them on the fly from the original data and compresses state-of-the-art proximity graph indices while preserving search accuracy. It reduces index size by up to 50x compared to conventional indices, maintains state-of-the-art accuracy, and achieves comparable latency. LEANN also supports storage-efficient index construction and updates, making it feasible to deploy vector search on personal devices or large-scale datasets.

Key Features

Recomputes embeddings on the fly instead of storing them, eliminating major storage costs
Compresses proximity graph indices while preserving search accuracy
Reduces index size by up to 50x compared to conventional indices
Maintains state-of-the-art accuracy for vector search
Supports storage-efficient index construction and incremental updates
Designed for deployment on personal devices and large-scale datasets

Pros & Cons

Pros
  • Dramatic storage savings (up to 50x reduction) compared to traditional indices
  • Eliminates need to store high-dimensional embeddings, freeing up memory
  • Maintains state-of-the-art search accuracy despite reduced storage
  • Supports practical index construction and dynamic updates
  • Enables private, device-local RAG without cloud dependencies
Cons
  • May incur higher computational cost for embedding recomputation at query time
  • Latency is comparable but not necessarily faster than traditional indices
  • Limited to applications where original data (e.g., text chunks) is available for recomputation

Best For

Retrieval-augmented generation (RAG) on personal devicesRecommendation systems with limited storage budgetsVector search in resource-constrained environmentsLarge-scale embedding-based search applications

Alternatives to LEANN

cognee

The memory for your AI Agents in 6 lines of code

databend

Data Agent Ready Warehouse : One for Analytics, Search, AI, Python Sandbox. — rebuilt from scratch. Unified architecture on your S3.

awesome-llm-apps

100+ AI Agent & RAG apps you can actually run — clone, customize, ship.

claude-mem

A Claude Code plugin that automatically captures everything Claude does during your coding sessions, compresses it with AI (using Claude's agent-sdk), and injects relevant context back into future ses

deeplake

Deeplake is AI Data Runtime for Agents. It provides serverless postgres with a multimodal datalake, enabling scalable retrieval and training.

anything-llm

The all-in-one AI productivity accelerator. On device and privacy first with no annoying setup or configuration.

FAQ

What problem does LEANN solve?
LEANN addresses the high storage overhead of vector indices, which can be several times larger than the original data, making deployment on personal devices or large datasets impractical.
How does LEANN achieve storage reduction?
It avoids storing embeddings by recomputing them on the fly from the original data and compresses proximity graph indices, reducing index size by up to 50x.
What applications can benefit from LEANN?
Applications such as retrieval-augmented generation (RAG), recommendation systems, and any embedding-based vector search that requires efficient storage.
Does LEANN compromise search accuracy?
No, LEANN maintains state-of-the-art accuracy while providing comparable latency for RAG applications.