Vector Retrieval with Compressed Candidate Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing vector retrieval solutions, particularly in shared storage scenarios, suffer from high delays due to synchronous disk I/O operations and network I/O costs, leading to increased latency and reduced user experience.

Innovation Solution

Store compressed vectors in a high-performance first-level storage resource and original vectors in a lower-performance second-level resource, using a pre-trained classification model to identify candidate vectors from the first-level resource, followed by similarity calculations on the second-level resource to reduce data access and volume.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Volume of stationary object

If original vectors are stored in a second-level storage resource with lower data access performance, then storage capacity is improved, but data access delay increases

Engineering Contradiction:
Improvestorage capacityVSAvoiddata access delay
Core Design Contradiction:
Volume of stationary objectVSLoss of time

Solution Approach 1:

The patent divides the storage system into two levels: a first-level storage resource for compressed vectors with high data access performance, and a second-level storage resource for original vectors with larger storage capacity but lower data access performance. This segmentation allows the system to leverage the advantages of both storage levels by performing preliminary similarity calculations on compressed vectors in the first level and only retrieving full original vectors for promising candidates to the second level.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If all original vectors are retrieved from the second-level storage resource for similarity calculation, then retrieval accuracy is improved, but data access volume and delay increase

Engineering Contradiction:
Improveretrieval accuracyVSAvoiddata access delay
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary similarity calculations using compressed vectors stored in the first-level storage resource before retrieving original vectors from the second-level storage resource. This preliminary action filters out most candidate vectors that are unlikely to be the most similar, so that only a small subset of promising candidates requires retrieval of the full original vectors for final accuracy verification.

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If compressed vectors are used for preliminary filtering, then data access volume is reduced, but retrieval accuracy may be compromised

Engineering Contradiction:
Improvedata access volumeVSAvoidretrieval accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent uses compressed vectors as an intermediary representation for preliminary similarity calculations. These compressed vectors serve as a mediator that enables fast filtering while the final retrieval accuracy is ensured by calculating similarity between the query vector and the original vectors for the filtered candidates. The compressed vectors do not directly determine the final result but facilitate the process by reducing the search space.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4685660A1Vector retrieval methods and apparatuses, devices, and storage media
Publication Date: 2026.01.28 ALIPAY (HANGZHOU) INFORMATION TECH CO LTD
  • EP4685660A1 patent drawingFigure 1~2
  • EP4685660A1 patent drawingFigure 3~4
  • EP4685660A1 patent drawingFigure 5

AI summary

One or more embodiments of this application provide a vector retrieval method and apparatus, a device, and a storage medium. A compressed vector corresponding to an original vector is stored in a first-level storage resource; the original vector is stored in a second-level storage resource; and the method includes: in response to a query vector, obtaining, from the first-level storage resource, at least one compressed vector having a highest similarity with the query vector; performing classification decision-making on each compressed vector based on a classification model, so as to obtain at least one target compressed vector whose corresponding original vector is a candidate vector most similar to the query vector; obtaining at least one target original vector corresponding to the at least one target compressed vector from the second-level storage resource; and respectively calculating a similarity between the query vector and the at least one target original vector, sorting the at least one target original vector based on the calculated similarity, and determining, based on the sorted at least one target original vector, a vector retrieval result corresponding to the query vector.