LLM Query Clustering for Distinguishing Similar User Requests
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLMs) face challenges in distinguishing between semantically similar queries with subtle differences and require frequent re-training due to platform changes, leading to outdated responses and inefficiencies in query resolution.
Innovation Solution
A system that clusters queries based on semantic similarity using LLMs, allowing for refined cluster creation and response generation by matching incoming queries to specific clusters, leveraging embeddings and domain-specific keywords.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If LLMs are used to generate query responses, then response generation capability is improved, but the ability to distinguish semantically similar queries deteriorates
Solution Approach 1:
The system segments queries into clusters based on semantic similarity before processing. By dividing the query space into distinct clusters, the system can handle semantically similar queries as groups rather than individual items, improving the ability to distinguish between different semantic categories while maintaining automated response generation.
Solution Approach 2:
The patent introduces an intermediary clustering mechanism between query input and LLM response generation. This intermediary layer processes queries through semantic clustering and representative selection, acting as a mediator that enhances query distinction before the LLM generates responses, thereby resolving the contradiction between automation and precision.
2Adaptability or versatility
If LLMs are re-trained frequently to adapt to platform changes, then response relevance is improved, but system efficiency deteriorates
Solution Approach 1:
The system performs preliminary clustering and identifies representative queries in advance, before actual query processing. By pre-organizing queries into clusters and selecting representatives, the system adapts to platform changes through structural organization rather than frequent re-training, maintaining response relevance while preserving system efficiency.
Solution Approach 2:
The patent uses representative queries that copy or represent entire clusters of semantically similar queries. Instead of re-training the LLM for every platform change, the system updates representative queries and cluster structures, which then propagate changes across all similar queries, maintaining adaptability without the overhead of frequent full re-training.
3Productivity
If queries are clustered by semantic similarity, then query processing efficiency is improved, but complexity of query analysis deteriorates
Solution Approach 1:
The system applies different processing qualities to different parts of the query space. Within each cluster, queries are treated as semantically equivalent and processed efficiently as a group. The clustering mechanism creates local regions of uniformity, allowing efficient batch processing while managing overall system complexity through localized handling.
Data Source
AI summary
A computer-implemented method is disclosed. The method includes: clustering a set of queries into first clusters; identifying, using a first large language model (LLM), queries in the first clusters that are semantically dissimilar from other queries in their cluster; clustering the queries identified as semantically dissimilar into one or more further clusters; matching a further query to a particular cluster from the first clusters or the further clusters; and obtaining a response to the further query based on providing, to a second LLM, data associated with the particular cluster.