LLM Document Chunk Generation With Hallucination Evaluation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The creation of textual documents, such as statements of work, is tedious and prone to errors due to the need for knowledge and experience to ensure coherence and avoidance of confusing or impermissible terms, especially when defining the goals and obligations of involved parties.

Innovation Solution

A method using a large language model to generate and evaluate document chunks based on topical information, with prompts and context, and an evaluation module to identify and address hallucinations, ensuring coherence and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a large language model generates document chunks independently, then generation speed increases, but coherence and consistency between chunks deteriorate

Engineering Contradiction:
Improvedocument generation speedVSAvoiddocument coherence
Core Design Contradiction:
ProductivityVSStability of the object's composition

Solution Approach 1:

The document generation process is divided into independent chunks that can be generated in parallel, with each chunk processed separately by the large language model while maintaining overall document coherence through structured prompting and evaluation

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

An evaluation module assesses generated chunks for hallucinations and coherence, providing feedback to identify and correct inconsistencies, ensuring the final document maintains stability and coherence across all generated sections

Inventive Principle:
Principle #23Feedback

2Reliability

If manual review and editing of document chunks is performed, then document accuracy improves, but time consumption increases

Engineering Contradiction:
Improvedocument accuracyVSAvoidtime consumption
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The evaluation module automatically detects and flags hallucinations in generated chunks without requiring manual review, enabling the system to self-correct errors through automated feedback mechanisms, thus maintaining high accuracy while minimizing time loss

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

Automated evaluation provides immediate feedback on generated chunks, identifying hallucinations and inconsistencies that can be corrected programmatically, replacing time-consuming manual review with efficient automated quality assurance

Inventive Principle:
Principle #23Feedback

3Reliability

If comprehensive evaluation for hallucinations is performed, then document reliability improves, but processing complexity increases

Engineering Contradiction:
Improvedocument reliabilityVSAvoidprocessing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The evaluation module acts as an intermediary between the generation and final assembly processes, systematically identifying and flagging hallucinations without requiring complex manual intervention, thus improving reliability while managing processing complexity through automated mediation

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentUS20250348686A1Document generation and evaluation using a large language model
Publication Date: 2025.11.13 INSIGHT DIRECT USA INC
  • US20250348686A1 patent drawing
  • US20250348686A1 patent drawing
  • US20250348686A1 patent drawing

AI summary

A method of generating a document having multiple chunks of text that collectively form the document is disclosed herein that includes receiving topical information; dependent upon the topical information, determining a first chunk of text to generate; retrieving at least one example first chunk of text; providing the at least one example first chunk of text and at least a portion of the topical information to a first large language model; prompting, by a computer processor, the first large language model to generate the first chunk of text through the use of a first request that includes a prompt that states a desired purpose of the first chunk of text to be generated, a context that provides information dependent upon the topical information, and the at least one example first chuck of text; and generating the first chunk of text dependent upon the topical information.