Intent Classifier Grounding for Generative Model Output
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional online search engines and generative models struggle with providing relevant and up-to-date information, especially for open-ended queries and intent-specific requests, leading to inefficient user interactions and excessive computational resources.
Innovation Solution
A workflow model that includes an intent classifier to generate prompts for a generative model, enhancing the output with supplemental content from data stores, thereby improving the relevance and timeliness of the information provided to users.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional search engines execute separate searches for each query refinement, then relevant results can be obtained, but computational resources are excessively consumed and response time increases
Solution Approach 1:
The system performs preliminary intent classification on the user's query before executing searches. The intent classifier analyzes the query to determine the user's information need category (e.g., navigational, informational, transactional), allowing the system to pre-filter and prioritize search strategies, thereby reducing unnecessary computational searches while maintaining result relevance.
Solution Approach 2:
The patent introduces an intent classifier as an intermediary component between the user query and the search execution engine. This mediator translates natural language queries into structured intent representations, which then guide the search process. This intermediary layer reduces computational overhead by enabling more efficient query routing and reducing the need for multiple iterative searches.
2Speed
If conventional generative models rely solely on training data, then they can generate content quickly, but the output becomes stale and not responsive to recent or intent-specific queries
Solution Approach 1:
The system merges the strengths of conventional search engines and generative models by combining intent-specific grounding information from structured data sources with the generative model's content creation capabilities. The intent classifier determines which data sources to query, and the generative model synthesizes this grounded information into natural language responses, achieving both timeliness and relevance.
Solution Approach 2:
The system performs preliminary intent classification and data source selection before generating the final response. By pre-determining the appropriate grounding information based on query intent, the system reduces the generative model's reliance on potentially outdated training data while maintaining fast response times through efficient information retrieval.
3Reliability
If users manually refine queries through multiple iterations, then relevant results can be obtained, but user time and interaction complexity increase
Solution Approach 1:
The system implements feedback loops where the intent classifier continuously monitors user interactions and refines its understanding of user intent based on subsequent queries and interactions. This feedback mechanism allows the system to adapt to user needs dynamically, reducing the number of refinement iterations required while maintaining high result relevance.
Solution Approach 2:
The patent enables the system to automatically perform query refinement and intent analysis without requiring manual user intervention. The intent classifier autonomously interprets user needs and adjusts search parameters, effectively allowing the system to serve itself in the query refinement process, thereby reducing user time investment while maintaining result quality.
4Ease of manufacture
If conventional systems use generic prompts for generative models, then implementation is simple, but output quality and intent-specific relevance deteriorate
Solution Approach 1:
The system applies local quality by tailoring prompts to specific intent categories rather than using uniform generic prompts. The intent classifier determines the appropriate prompt template and grounding information based on the detected intent type, ensuring that each prompt is optimized for its specific purpose (e.g., product recommendations, informational queries, navigational requests), thereby significantly improving output quality while maintaining systematic simplicity.
Data Source
AI summary
A computing system for generating enhanced output of a generative model is disclosed. The computing system is configured to receive an input set forth by a user of a client computing device that is in network communication with the computing system. An intent classifier produces, from the input, an output indicative of a user intent. The output of the intent classifier and the input are provided as input into a workflow model. The workflow model generates a prompt based upon the input and the user intent and provides the prompt as input into a generative model, which causes the generative model to generate an output based upon the prompt. The workflow model receives the output of the generative model and identifies supplemental content related to the output. The workflow model then generates an enhanced output based upon the output and the supplemental content for presentation to the user.