LLM-Generated Expressions for Secure Local UI Data Modification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data entry and manipulation methods in software applications, such as ERP systems, rely on user input and are hindered by security risks, data communication restrictions, and substandard performance of large language models (LLMs) in mathematical operations.
Innovation Solution
Use a bidirectional encoder model to generate expressions for data manipulation by sending identifiers and data types to the LLM, rather than actual data, and parse these expressions to modify the data locally, enhancing security and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If actual data is sent to the LLM for data manipulation, then the LLM can understand and process the data context, but security risks increase due to exposure of sensitive data
Solution Approach 1:
The patent introduces an intermediary process that transforms actual data into synthetic data with preserved structural characteristics. This intermediary transformation layer allows the LLM to process data-like information without direct exposure to sensitive actual data, thus maintaining security while enabling effective data manipulation.
Solution Approach 2:
The patent creates synthetic copies of data structures and patterns that mimic the characteristics of actual data without containing the actual sensitive information. These copies are used for LLM processing, allowing the model to learn and manipulate data patterns while the original sensitive data remains protected on-premises.
2Adaptability or versatility
If large amounts of data are sent to the LLM for processing, then comprehensive data manipulation is possible, but data communication restrictions and input size limits are exceeded
Solution Approach 1:
The patent extracts only the essential structural characteristics and patterns from large datasets, separating these from the bulk data volume. By extracting and transmitting only the necessary structural information in synthetic form, the system enables comprehensive data manipulation capability while staying within communication and input size constraints.
Solution Approach 2:
The patent segments the data processing task into multiple components: local data preparation, synthetic data generation, LLM processing of condensed information, and result application. This segmentation allows the system to handle large data volumes by distributing processing across multiple stages rather than requiring all data to be transmitted at once.
3Ease of operation
If LLM is used for mathematical operations, then language-based data manipulation is possible, but accuracy deteriorates due to substandard mathematical performance
Solution Approach 1:
The patent performs preliminary preparation of data in synthetic form with explicit structural and mathematical characteristics before LLM processing. By pre-structuring the data with clear mathematical relationships in the synthetic representation, the LLM can more accurately perform mathematical operations without relying solely on its inherent mathematical capabilities.
Solution Approach 2:
The patent transforms data into a synthetic parameter space where mathematical relationships are explicitly represented through structured parameters rather than raw values. This parameter transformation allows the LLM to operate more effectively on mathematical relationships by working with explicitly defined structural parameters instead of implicit numerical data.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In an example embodiment, rather than send the state of the UI screen as a whole to an LLM to generate or modify one or more values on the UI screen, the identifiers and data types of data on the UI screen are gathered and sent to the LLM. The LLM is instructed to return a list of expressions that define what needs to be done to satisfy a user prompt. The calling process then evaluates the list of expressions to actually manipulate the data content.