Function Description

The Ascend-developed vector retrieval algorithm enables approximate retrieval in a high-dimensional, large base library. It uses a proprietary matrix approximation policy to compress feature vectors and store them in the base library. Lastly, a custom retrieval policy is employed to find the topK most approximate vectors within the base library.

The vectors stored in a base library and the query vectors of each API must be of the normalized float type.

Multi-thread concurrent calling is not supported. Therefore, you need to add locks in multi-thread scenarios. Otherwise, the retrieval API may be abnormal. In addition, one device cannot be shared by different threads.

This algorithm is primarily used for approximate fuzzy search in large-scale base libraries. Compared with brute-force search, its precision is reduced. In the small base library scenario, you are advised to increase the hyperparameter value to minimize precision loss.