Vector Search Filter Directions for Efficient Query Redirection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vector-based search techniques struggle with efficiently applying additional criteria to filter search results without negatively impacting performance time, accuracy, or requiring significant additional storage capacity, especially when using pre-processing or post-processing techniques.
Innovation Solution
Implement filter direction modifications that utilize a filter matrix to guide vector-based searches to relevant subspaces, approximating relevant filter directions and applying them to query vectors without altering the underlying search technique or rebuilding indexes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If pre-processing or post-processing techniques are used to apply filter criteria to vector-based search results, then filtering capability is improved, but search performance time deteriorates and computational overhead increases
Solution Approach 1:
The system pre-computes filter directions and stores them in a filter matrix during an offline phase. This preliminary action allows the filter criteria to be ready and integrated into the search process without adding computational overhead during actual search operations, thus improving filtering capability while maintaining search performance time
Solution Approach 2:
The patent introduces a filter matrix as an intermediary structure that bridges the query vector and the filter criteria. The filter matrix contains pre-computed directions that guide the query through the filter space, enabling efficient filtering without requiring separate pre-processing or post-processing steps, thereby maintaining productivity while improving adaptability
2Measurement precision
If filter criteria are applied to vector-based search results, then search accuracy is improved, but storage capacity requirements increase
Solution Approach 1:
The system changes the parameter representation of filter criteria from storing full filter definitions to storing compact filter directions in a matrix structure. Each filter is represented as a direction vector in the same space as the query vectors, allowing accurate filtering while using minimal additional storage capacity proportional to the number of filters times the vector dimension
3Manufacturing precision
If additional filter criteria are applied to search results, then filtering precision is improved, but computational overhead increases
Solution Approach 1:
Filter directions are pre-computed and stored in a matrix structure during an offline phase, eliminating the need for computationally intensive filter calculations during actual search operations. This preliminary computation reduces computational overhead to simple matrix-vector operations while maintaining high filtering precision
Solution Approach 2:
The patent replaces complex filter application mechanics with simple vector arithmetic operations. Instead of applying filters as separate processing steps, the system uses matrix-vector multiplication to integrate filter criteria directly into the search process, substituting heavy computational mechanics with efficient linear algebra operations
Data Source
AI summary
Filter direction modifications are performed in vector-based searches. A query data item is obtained along with a data item filter to perform a data item search on a set of data items. A query vector that represents the query in feature space is generated and modified using a filter vector obtained from a matrix of filter vectors generated for the data item filters. The modified query vector is used to perform a search technique to identify data items to return as a result for the search.