@XAMTO_AI: Traditional vector stores save all embeddings, easily reaching hundreds of GB for millions of documents. LEANN switches to a simplified graph, computing vectors on the fly for only the nodes traversed during retrieval. In the paper and README, large-scale comparisons show about 97% savings, with recall still close to full HNSW. Suitable for searching local files, emails, browsing history, and … on your own laptop.
Summary
LEANN is a new vector indexing method that reduces storage requirements by about 97% through graph simplification and real-time embedding computation, while maintaining retrieval recall close to full HNSW. It is suitable for use on laptops and won the Best Paper award at MLSys 2026.
View Cached Full Text
Cached at: 09/24/26, 08:30 PM
We track zero telemetry. This survey is the ONLY way to tell us if you want GPU Acceleration or More Integrations next. 👉 Click here to cast your vote (2 mins)
💬 Join our Slack community!
We’d love for you to be part of the LEANN community! 👉 Join LEANN Slack If the invite link has expired or you have trouble joining, please open an issue and we’ll help you get in!
The smallest vector index in the world. RAG Everything with LEANN!
⭐ Star us on GitHub if Leann is useful for your research or applications!
Made with ❤️ by the Leann team
Similar Articles
@vintcessun: Compressing 10 million vectors from 31GB to 4GB, with search even faster than FAISS — sounds crazy, but Turbovec actually did it. The core is Google's TurboQuant data-independent quantization: no training, no parameter tuning, just add vectors and index. Handwritten NEON/AVX-512 implementations are genuinely 12-20% faster, supporting filtered search by ID, saving a ton of post-processing hassle. Rust under the hood + pip install, minimal maintenance cost.
Turbovec, based on Google's TurboQuant algorithm, compresses 10 million vectors from 31GB to 4GB, with search speed 12-20% faster than FAISS, supports filtered search, and offers a Rust implementation with a Python package.
@Ryrenz: Incredible, packs the entire AI memory system into a single file. GitHub has reached legendary status with 16.2K stars. To build a knowledge base for AI, you typically need: a vector database service, a full-text search service, and store metadata in another database. During development, run Docker Compose, and during deployment, configure three components...
Memvid is a tool that compresses the AI memory system into a single file, supporting vector search and full-text retrieval, written in Rust, with high performance and portability.
@vista8: Last night I casually tested Knowly developed by the @Ethan_Yang_AI team. Tried interpreting YouTube videos and arXiv papers – the results were stunning. Except for a rather limited free quota and slightly slow vector processing. In terms of both product interaction and interpretation quality, it's no less impressive than NotebookLM. With a Chrome extension that has only a few users but has already been selected by Google as a featured pick, its strength is evident. Official site in comments https://t.co/62NkT3pO4G
Introduces Knowly AI tool, capable of interpreting YouTube videos and arXiv papers with impressive results. Interaction and interpretation quality rival NotebookLM. Comes with a Chrome extension already featured by Google. Drawbacks: limited free quota and slightly slow vector processing.
@nash_su: http://WisMe.ai Knowledge Graph Feature Update. The old version had poor readability and performance issues when content became too large. After the update, it remains smooth even with hundreds or thousands of nodes. Meanwhile, the clustering function is more practical, and it also calls LLM for information summarization, making search and browsing more convenient and efficient. Harness …
WisMe.ai updates its knowledge graph functionality for better performance with hundreds to thousands of nodes, improved clustering, and LLM-powered summarization for more efficient search and browsing.
@vintcessun: Feeding too many documents into RAG causes retrieval quality to drop from 75% to 40%? Vector search is diluted by a large amount of irrelevant content, causing a sharp drop in hit rate in real deployment. Root cause: heterogeneous documents are retrieved together, noise drowns out signal. Multi-agent orchestration seems intelligent but actually introduces a precision-fidelity paradox—poor configuration leads to failure in both aspects. The paper proposes MA…
This paper identifies 'vector search dilution' in RAG systems when scaling to large heterogeneous document collections, where accuracy dropped from 75% to 40% in a real-world deployment. The proposed MASDR-RAG method uses domain scoping via organizational metadata before retrieval, improving P@10 from 0.77 to 0.86 with low cost and easy deployment.