Iterative Vector Retrieval for Stable Enterprise Query Results
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models lack access to confidential or unpublished information, necessitating a method to combine publicly available data with enterprise-specific data for comprehensive query responses.
Innovation Solution
Implementing retrieval-augmented generation by using an iterative data retriever that searches databases based on query vectors to generate result sets, combining them with large language models to include both public and enterprise data, and employing an iterative approach to maintain stable and efficient result cardinality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single database search is performed for query retrieval, then the retrieval speed is fast, but the comprehensiveness of results (including both public and enterprise data) is insufficient
Solution Approach 1:
The patent divides the retrieval system into two separate vector databases: a first vector database storing public data and a second vector database storing enterprise-specific confidential data. This segmentation allows independent optimization of each database while ensuring both public and enterprise information are retrieved comprehensively, resolving the contradiction between retrieval speed and result completeness.
Solution Approach 2:
The patent merges results from multiple database searches by combining the first result set (from public data) and the second result set (from enterprise data) into a unified third result set. This combination ensures comprehensive information retrieval while maintaining efficient parallel processing of both databases, thus achieving both speed and completeness.
2Loss of information
If iterative retrieval is performed to improve result quality, then the comprehensiveness of results improves, but the computational complexity and processing time increase
Solution Approach 1:
The patent implements an iterative retrieval process where the quality metric of the combined result set is evaluated against a threshold. If the quality metric falls below the threshold, the system performs additional iterative retrievals with modified parameters. This feedback mechanism ensures high-quality comprehensive results while avoiding unnecessary iterations when the threshold is already met, thus balancing completeness with computational efficiency.
Solution Approach 2:
The patent performs retrieval operations that may initially exceed the minimum required effort by searching both public and enterprise databases with potentially overlapping query parameters. However, the system optimizes by combining results and applying quality metrics to filter and refine the final output, ensuring comprehensive coverage without excessive computational waste.
3Loss of information
If multiple database searches are conducted in parallel, then the comprehensiveness of results improves, but the resource consumption and processing overhead increase
Solution Approach 1:
The patent segments the computational workload by dividing it into two independent parallel search operations: one querying the public data database and another querying the enterprise data database. Each database search is optimized independently, allowing efficient resource utilization while ensuring both public and enterprise information are retrieved simultaneously, thus achieving comprehensive results without excessive resource consumption.
Solution Approach 2:
The patent merges the results from parallel database searches into a unified result set, eliminating redundant processing. By combining the first result set and second result set and applying quality metrics, the system achieves comprehensive information retrieval while optimizing resource usage through intelligent result consolidation rather than processing all possible combinations.
Data Source
AI summary
A system and methods for iterative data retrieval. In some embodiments, a method includes: obtaining a first result set, from a first vector database, the obtaining being based on proximity to a first vector, the first vector being associated with a query; determining that an iteration criterion is met, the iteration criterion being based on the first result set; and obtaining a second result set from the first vector database, based on the determining that the iteration criterion is met.