UC Berkeley introduced LEANN — an open-vector index for RAG systems. According to MarkTechPost, it stores a pruned graph, and it re-computes embeddings during query processing.
In the described scenario, the index for 60 million fragments reduces the size from 201 to 6. It also claims an HNSW-like recall of 90% at 4 instead of 188, but the search takes seconds.
The value of the approach lies in the trade-off between storage and response speed. The available materials are presented by a single independent source and are metadata-only, so full testing conditions and reproducibility require verification.
editorial commentary
Why it matters
A likely consequence is interest in indexes that save space at the cost of response delay, especially in scenarios with modest speed requirements. The next verifiable signals will be the publication of code, testing methodology, and independent comparisons. A significant uncertainty is how the claimed results hold on other datasets and loads.