Conventions

Mx/mx

-

MindX (general name of Ascend application enablement software).

Ascend

-

Ascend (Ascend application enablement software)

Feature retrieval (FeatureRetrieval/AscendFaiss)

Heterogeneous retrieval component developed based on Faiss on the Ascend platform.

Faiss

Open-source search framework of Facebook

Protobuf

Open-source data serialization framework of Google

OpenBLAS

Open-source Basic Linear Algebra Subprograms (BLAS) and LAPACK API mathematical calculation package

Index

A group of vectors encapsulated by AscendFaiss/Faiss. It is used to manage the group of vectors and perform efficient retrieval.

Small library retrieval/Exact search

For global search, the base library contains 300,000 to 1,000,000 records. Currently, brute-force search algorithms, such as Flat, SQ, and INT8, are supported.

Large library retrieval/Inexact search

This function is developed based on IVF. The library contains more than ten millions of records. Currently, such algorithms as IVFPQ, IVFSQ, and IVFINT8 are supported.

Feature vector query

After the user enters the feature vector to be queried, the small and large library retrievals calculate the distance between the feature vector to be queried and all feature vectors in the base library, and return the top K results with the highest similarity to complete one feature vector retrieval.

Library creation

Add the feature library with millions or tens of millions of records to the Ascend platform.

Base library addition

Add new feature data to the Ascend platform based on the existing base library.

Base library deletion

Delete the feature vector based on a specified index from the base library. Multiple feature vectors can be deleted at a time.

Base library storage/flushing

Save the features and indexes of the user base library locally. Therefore, you do not need to train or create a library when restoring services.

Base library restoration/loading

Restore the saved feature library and indexes to the Ascend platform.