Systems and methods for information retrieval

Inventors

SHRESTHA, Susav Lal • Li, Zongwang • ANNAPAREDDY, Narasimha

Assignees

Samsung Electronics Co Ltd • Texas A&M University System

Abstract

Provided are systems, methods, and apparatuses for systems and methods of memory efficient multi-vector information retrieval based on embeddings from a storage pipelined network. In one or more examples, the systems, devices, and methods include performing a first portion of a nearest neighbor search on a first subset of a set of media sources; performing a fetch process on the first subset that include identifying a first set of highest matching media sources from the first subset, transferring multi-vector representations of the first set of highest matching media sources from a storage drive to a memory, and performing a second ranking of the first set highest matching media sources. The systems, devices, and methods include performing, in parallel with the fetch process, a second portion of the nearest neighbor search on a second subset of the set of media sources.

Core Innovation

The invention relates to a memory-efficient multi-vector information retrieval approach using “Embedding from Storage Pipelined Network (ESPN)”. It performs a nearest neighbor search by comparing a single vector representation of a query to single vector representations of media sources, split into a first subset and a second subset. A first portion of the nearest neighbor search is performed on the first subset, while a second portion of the nearest neighbor search is performed in parallel on the second subset based on the same single vector representation of the query.

During the fetch process on the first subset, a first set of highest matching media sources is identified from the first subset based on a first ranking of the first subset determined from the first portion of the nearest neighbor search. Then multi-vector representations associated with the first set of highest matching media sources are transferred from a storage drive to a memory based on the first ranking. After the transfer, a second ranking is performed on the first set of highest matching media sources.

The invention further addresses cases where top results are missed by detecting a missing media source via identifier comparison between a first set of highest matching media sources and a second set of highest matching media sources. The multi-vector representation of the missing media source is transferred from the storage drive to the memory. Optionally, a third set of highest matching media sources is determined by performing a third ranking based on the first set, the second set, and the missing media source.

Claims Coverage

The independent claim set covers three forms: a method, a device, and a non-transitory computer-readable medium. Across these, the core coverage includes four main inventive feature groups: split nearest neighbor search on first and second subsets using a single vector query representation, ranked fetch of multi-vector representations from a storage drive to memory for the first subset’s highest matching media sources, parallel execution of the second subset search with the fetch process, and missing-media-source identification using identifier comparison followed by an additional ranking.

Collectively, the independent claims require a two-subset nearest neighbor search using a single vector representation of the query, a ranked fetch of multi-vector representations from a storage drive into memory for the top matches from the first subset, parallel execution with a second subset search, and an optional missing-media-source identifier comparison followed by an additional ranking.

Stated Advantages

Documented Applications