摘要 |
<p>Similar vectors are fast retrieved from a vector database of several hundreds of dimensions with reference to a single vector index according to the criterion of the inner product or distance, after specifying the similarity search range and the maximum number of similar vectors to be retrieved. For creating the vector index, each vector is decomposed into sub-vectors and featured by a norm section, an assigned area, and an argument section. For similarity search, a sub-query vector and a sub search range are determined from the query vector and the search range, similarity search in sub-space is carried out, and differences from the search range are cumulated to determine the upper limits. An accurate criterion having a higher upper limit is preferentially determined, thereby producing a final similarity search result.</p> |