Enterprise LLM Operation Assistance Through Tabular Data Serialization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Integrating large language models (LLMs) into enterprise operations is challenging due to inaccurate and misleading outputs, inability to comprehend structured tabular data, lack of domain-specific context, and limited mathematical computation capabilities, which hampers their effectiveness in tasks like employee compensation reviews.
Innovation Solution
Implementing tabular data pre-processing to serialize data and using prompt engineering techniques such as chain-of-thought (CoT) and reasoning-acting (ReAct) to inject persona and thinking style, enabling LLMs to accurately analyze and compute for enterprise operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If LLMs are directly integrated into enterprise operations, then automation capability is improved, but output accuracy deteriorates due to inaccurate and misleading outputs
Solution Approach 1:
The patent introduces an intermediary processing layer between the LLM and enterprise operations that includes: (1) A prompt engineering module that structures inputs with domain-specific context, (2) A output validation module that verifies LLM outputs against enterprise rules and data consistency, and (3) A feedback loop that refines future prompts based on output quality. This intermediary layer enables automation while ensuring reliability through multiple checkpoints.
2Speed
If LLMs are used for enterprise operations, then task processing speed is improved, but data comprehension capability deteriorates due to inability to comprehend structured tabular data
Solution Approach 1:
The patent segments the data processing pipeline into distinct modules: (1) A tabular data parser that converts structured tables into structured query language or JSON formats, (2) A context extraction module that identifies relevant relationships and constraints from the tabular data, and (3) An LLM processing module that receives pre-processed data in a comprehensible format. This segmentation allows the LLM to process data quickly while maintaining comprehension accuracy through specialized preprocessing.
3Productivity
If LLMs are integrated into enterprise operations, then operational efficiency is improved, but domain-specific context capability deteriorates due to lack of domain-specific context
Solution Approach 1:
The patent implements preliminary action by: (1) Pre-processing enterprise data into standardized formats with embedded domain context before LLM processing, (2) Pre-configuring prompt templates with domain-specific instructions and constraints, and (3) Pre-training or fine-tuning the LLM on enterprise-specific data sets. These preliminary actions ensure the LLM operates efficiently with appropriate domain context already in place, eliminating the need for real-time context adaptation.
4Extent of automation
If LLMs are used for enterprise operations, then task automation is improved, but mathematical computation capability deteriorates due to limited mathematical computation capabilities
Solution Approach 1:
The patent introduces a mathematical computation intermediary layer that: (1) Identifies mathematical operations required in the task, (2) Translates LLM-generated mathematical expressions into executable code or uses a code interpreter, and (3) Validates computational results against expected ranges and constraints. This intermediary ensures accurate mathematical computation while maintaining task automation.
Data Source
AI summary
Methods, systems, and computer-readable storage media for receiving tabular data, serializing the tabular data to provide serialized data, generating a prompt comprising a persona, a set of chain-of-thought (CoT) steps, and a thinking style, the persona being specific to an operation of the enterprise and including a natural language description of a role for executing the operation, the CoT steps defining a sequence of actions that a LLM is to perform in processing the prompt, the thinking style including a natural language description of how the LLM is to process the prompt, transmitting the prompt and serialized data to the LLM, receiving output of the LLM responsive to the prompt, and executing at least one operation using the output.