LLM Content Verification Using RAG and ER Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI-generated content often suffers from hallucination, producing false, misleading, or illogical information due to reliance on incomplete, outdated, or low-quality training data, which undermines its reliability, especially in critical applications.
Innovation Solution
A layered verification system comprising a content generation LLM and a verifier LLM, utilizing Retrieval Augmented Generation sources and an entity, property, and relationship database to verify content accuracy by comparing against external dynamic knowledge bases, with the verifier LLM employing pre-training, pre-configuration, and optimization strategies for efficient verification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If AI generates content using static internal knowledge, then generation speed is improved, but content accuracy deteriorates due to hallucination
Solution Approach 1:
The patent introduces an external dynamic knowledge base as an intermediary between the content generation LLM and the verification process. The knowledge base acts as a mediator that provides real-time, accurate information to verify generated content, resolving the contradiction between fast generation using static knowledge and accurate verification requiring dynamic knowledge.
Solution Approach 2:
The verification system is segmented into multiple independent components: content generation LLM, external dynamic knowledge base, and verifier LLM. This segmentation allows the generation process to operate independently at high speed while the verification process separately validates accuracy against dynamic knowledge, resolving the speed-accuracy tradeoff.
2Device complexity
If AI relies on incomplete or outdated training data, then model complexity is reduced, but content reliability deteriorates due to hallucination
Solution Approach 1:
The system performs preliminary action by pre-training the verifier LLM on domain-specific data and pre-configuring the external dynamic knowledge base with comprehensive, up-to-date information. This preliminary preparation enables the verifier to effectively detect hallucinations without requiring the content generation LLM to be retrained with more complex or updated data.
Solution Approach 2:
The external dynamic knowledge base serves as an intermediary that compensates for the limitations of static training data. It provides current, accurate information that the verifier LLM can use to identify hallucinations, allowing the system to maintain simple model architecture while achieving high reliability through external verification.
3Measurement precision
If a verification system compares generated content against external dynamic knowledge bases, then content accuracy is improved, but system complexity increases
Solution Approach 1:
The patent extracts the verification function into a separate, specialized verifier LLM that operates independently from the content generation LLM. This extraction allows the verification system to focus solely on accuracy checking against the external knowledge base, improving measurement precision while managing complexity through functional separation.
Solution Approach 2:
The verifier LLM is designed with multi-functionality, capable of performing various verification tasks including factual accuracy checking, consistency validation, and hallucination detection. This universal verification capability improves content accuracy across different types of generated content without proportionally increasing system complexity.
Data Source
AI summary
A system to verify correctness of content is disclosed. The system may include one or more processors and a memory. The processors may obtain, by a content generation LLM, a user prompt via a user interface rendered on a user device, and generate a response to the user prompt responsive to obtaining the user prompt. The content generation LLM may be paired with Retrieval Augmented Generation (RAG) sources. The processors may transmit, by the content generation LLM, the response to a verifier LLM. The processors may parse, by the verifier LLM, the response into structured data, and compare the structured data with data stored in an entity, property, and relationship (ER) database that is paired with the RAG sources and an external database. The processors may determine, by the verifier LLM, correctness of response based on the comparison, and output the correctness of the response on the user interface.