Researchers at UC Berkeley have unveiled LEANN, an innovative open-source vector index designed to optimize the storage and retrieval of information in large-scale applications. This new index compresses a 60 million chunk Retrieval-Augmented Generation (RAG) index from 201 GB down to just 6 GB, representing a substantial reduction in data footprint.
LEANN achieves this remarkable efficiency by utilizing a pruned graph structure and recalculating embeddings at query time. It boasts an impressive accuracy rate, matching the performance of the Hierarchical Navigable Small World (HNSW) algorithm with a 90% recall rate, while operating with only 4 GB of memory as opposed to the 188 GB typically required. However, this efficiency comes with a trade-off, as the search times may extend to seconds.
This development is poised to have significant implications for the field of AI, especially in applications where rapid data retrieval and efficient storage are paramount. By reducing resource requirements, LEANN could enable broader access to advanced AI capabilities, particularly in environments with limited computational power.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.
