LLM Fact Auditor: Building a Trustworthy Fact-Checking Pipeline for Large Language Model Outputs

A post-processing pipeline for verifying the factual accuracy of content generated by large language models, integrating entity linking, answer extraction, and knowledge base cross-validation functions

• Original Authors/Maintainers: Joel Dettinger, Ruida Zhou, Hongqian Xia, Angelo De Nadai (VU Amsterdam Web Data Processing Systems Course Project Team)
• Source Platform: GitHub
• Original Title: LLM-Fact-Auditor
• Original Link: https://github.com/dettinjo/LLM-Fact-Auditor
• Publication Date: June 2, 2026

Large Language Models (LLMs) such as GPT and Llama have demonstrated amazing text generation capabilities, but they have a well-known problem—"hallucination". The model may confidently generate information that seems reasonable but is actually completely wrong. This poses a serious challenge in application scenarios requiring high accuracy (e.g., medical consultation, legal advice, news reporting).

Traditional solutions include using more powerful models, increasing training data, or fine-tuning, but these methods are costly and cannot completely eliminate hallucinations. The research team at VU Amsterdam took a different approach: since it is impossible to prevent the model from hallucinating, build a "fact audit" layer for the model output to verify the accuracy of the generated content through external knowledge bases.

LLM Fact Auditor is a post-processing pipeline that takes user questions and original LLM-generated answers, and outputs verified and entity-linked structured results through multi-stage processing. Developed as a practical project for the Web Data Processing Systems course, this project demonstrates how to combine natural language processing technology with knowledge graphs to build a practical fact-checking system.

The system first identifies named entities (people, places, organizations, etc.) in the text and links them to corresponding Wikipedia pages. This step provides the foundation for subsequent fact-checking—mapping mentions in free text to structured knowledge base entities.

For example, when the model answers "Managua is the capital of Nicaragua", the system will identify the two entities "Managua" and "Nicaragua" and link them to their respective Wikipedia pages.

LLM outputs are often verbose and contain a lot of explanatory content. The answer extraction module distills complex answers into concise direct answers, such as "yes/no" or specific entity names. This makes subsequent fact-checking more precise and efficient.

This is the core function of the system. The extracted answers are cross-validated with structured knowledge from Wikidata and content from linked Wikipedia pages to determine the correctness of the answers. The system outputs marks such as "correct", "incorrect", or those requiring further review.

The project uses a modern NLP technology stack:

• Python: Main development language
• Docker: Provides a standardized runtime environment
• Llama 3/Meta AI: Supports locally deployed LLM inference
• PyTorch & Hugging Face Transformers: Deep learning model support
• spaCy: Basic natural language processing tool
• Wikidata: Structured knowledge base

This technology selection reflects the principle of pragmatism—it not only leverages powerful pre-trained models but also reduces reliance on external APIs and costs through local deployment (Llama).