Show HN: Two-line change, 30% RAG boost

1 points by snasan 2 days ago · 0 comments · 1 min read

While graph-based vector retrieval algorithms like HNSW and NSG are highly favored for their excellent retrieval speed, it is often overlooked that they are designed for Euclidean space vector retrieval. "Metric mismatch" can be devastating in many scenarios—vector data better suited for maximum inner product retrieval, when paired with Euclidean vector algorithms, often leads to "retrieval results semantically irrelevant to the query." Looking back at the field of maximum inner product retrieval, there has yet to be a groundbreaking algorithm comparable to HNSW or NSG. Many previous works performed well only on certain datasets but suffered severe degradation when applied to others. To address this, in collaboration with Zhejiang University, we have successfully elevated maximum inner product retrieval research to the same level as Euclidean space retrieval. Our innovative method—PSP—can boost maximum inner product retrieval efficiency by 30% with minimal modifications to existing graph algorithms!

No comments yet.