Agent-Based Semantic Match Workflow for Nuanced Mortgage Queries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer-based mortgage systems struggle with flexible and nuanced mortgage product and rate data searches due to rigid rules-based decision trees, stochastic large language models, and overwhelming data sets, leading to unpredictable and inconsistent results.
Innovation Solution
An agent-based AI system with intent classification and multiple specialized large language models, employing a predefined workflow and action handlers to orchestrate operations, providing a dynamic and context-aware response through a tournament-style semantic match on tabular mortgage data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If rigid rules-based decision trees are used in mortgage systems, then system structure is simple and easy to implement, but the system lacks flexibility to process nuanced queries and dynamic mortgage offerings
Solution Approach 1:
The patent segments the mortgage system into multiple specialized agents (semantic match agent, normalization tagger agent, column selector agent) that each handle specific tasks. This segmentation allows the system to process nuanced queries flexibly while maintaining manageable complexity through modular design, where each agent can be independently optimized and maintained.
Solution Approach 2:
The patent introduces an action handler as an intermediary component that coordinates between the user query and the specialized agents. This intermediary manages the complexity by providing a structured interface that routes queries appropriately, allowing the system to handle diverse mortgage offerings and nuanced queries without requiring a complete rewrite of the entire system architecture.
2Ease of operation
If large language models are used autonomously to handle complex mortgage operations, then the system can process natural language queries, but the results become unpredictable due to stochastic nature of LLMs
Solution Approach 1:
The patent divides the autonomous LLM task into segmented roles with specific responsibilities (semantic matching, normalization, column selection). Each specialized agent focuses on a narrow task, which reduces the stochastic variability inherent in general-purpose LLMs and improves result predictability while maintaining natural language processing capabilities.
Solution Approach 2:
The patent changes the operational parameters of large language models by fine-tuning them for specific mortgage domain tasks and constraining their output to structured formats. This parameter adjustment transforms the stochastic nature of general LLMs into more predictable, task-specific responses that maintain ease of operation with natural language while improving reliability.
3Quantity of substance
If complete mortgage product records are presented to AI models, then comprehensive data is available for analysis, but the models become overwhelmed and unable to maintain attention and context
Solution Approach 1:
The patent extracts only the relevant columns and data fields needed for each specific mortgage query using the column selector agent. Instead of presenting complete mortgage product records to AI models, the system extracts and provides only the necessary subset of data, which maintains data completeness for the task at hand while preventing model overload and maintaining contextual attention.
Solution Approach 2:
The patent segments the large mortgage data set into manageable components through the column selector agent, which identifies and separates relevant fields from the complete product records. This segmentation allows comprehensive mortgage data to be processed effectively by breaking it down into focused, task-relevant subsets that AI models can handle without becoming overwhelmed.
4Productivity
If traditional systems use rigid filters or algorithms for exact matches, then processing is efficient and fast, but the systems struggle with nuanced queries that require logical leaps and contextual understanding
Solution Approach 1:
The patent changes the matching parameter from exact match to semantic similarity matching. The semantic match agent uses large language models to evaluate contextual relevance and semantic similarity rather than requiring exact matches, which enables the system to handle nuanced queries with logical leaps while maintaining improved processing efficiency through optimized similarity calculations.
Solution Approach 2:
The patent substitutes mechanical rigid filtering algorithms with AI-based semantic matching. Instead of using traditional exact-match filters that cannot handle nuanced queries, the system employs large language models with fine-tuned semantic understanding, replacing the mechanical filtering process with intelligent, context-aware matching that maintains productivity while significantly improving adaptability.
Data Source
AI summary
A method of performing a tournament process using a large language model included in an agent to process candidate records in a series of rounds, the candidate records selected by the third agent because they meet requirements and objectives identified by the at least one of the first tag and the second tag, the tournament process including acts of: randomly organizing the candidate records that resolve the mortgage query into subgroups that include a total number of candidate records less than a threshold quantity of candidate records; individually evaluating the candidate records in the subgroup to rank the candidate records within the subgroup based on how effectively the candidate records meet the requirement and the objective; and repeating in a series of rounds to identify top candidates to meet both the requirement concerning the eligible mortgage and the objective concerning the eligible mortgage.