Hallucination Detection for Large Models in Healthcare: A Comparative Evaluation Framework of RAG vs. Non-RAG Based on LangGraph

A hallucination evaluation project for large language models focused on medical Q&A scenarios, which quantifies the accuracy and hallucination rate of models in medical knowledge Q&A by comparing RAG-enhanced and pure generation modes.

Large language models are increasingly used in the healthcare field, but the hallucination problem remains a key obstacle to their practical deployment. When models generate medical information that seems reasonable but is inconsistent with facts, it may pose serious safety risks. This project focuses on medical Q&A scenarios and builds a systematic evaluation framework to quantitatively compare the hallucination performance of models under different configurations.

The project uses a streamlined and efficient tech stack:

• Orchestration Layer: LangGraph handles workflow orchestration
• Vector Storage: FAISS as the knowledge base retrieval backend
• Embedding Model: nomic-embed-text provided by Ollama
• Generation Model: llama3:latest deployed locally via Ollama

This architectural choice reflects the principle of pragmatism—achieving a complete RAG (Retrieval-Augmented Generation) pipeline without relying on external APIs.

The core of the project lies in comparing two working modes:

The model answers questions directly based on parametric knowledge, testing its inherent medical knowledge reserve and hallucination tendency. This mode reflects the baseline performance of general large models without optimization.

After retrieving relevant medical knowledge fragments via FAISS, the model generates answers. This mode evaluates whether retrieval augmentation can effectively suppress hallucinations and whether the introduced retrieval noise will bring new types of errors.

The project establishes multi-dimensional evaluation metrics:

1. Accuracy: Consistency between the answer and the standard answer
2. Error Rate: Proportion of obvious factual errors
3. Hallucination Categories: Fine-grained classification of hallucination types

In addition, the system is equipped with a verifier_agent to perform secondary verification on the generated results, forming a closed-loop evaluation mechanism of "generation-verification".

The project uses JSON format to maintain the medical knowledge base (data/knowledge_base.json) and supports rebuilding the FAISS index via command-line parameters. This design makes the update and maintenance of the knowledge base relatively flexible, facilitating customized expansion for specific medical fields (such as internal medicine, pharmacy).