KRAGEN: a knowledge graph-enhanced RAG framework for biomedical problem solving using large language models

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Knowledge Representation

Background:

• Large language models (LLMs) face challenges in factual consistency and data accuracy, particularly in specialized domains like biomedicine.
• LLM limitations include generating hallucinations and being influenced by noisy data, impacting trustworthiness and compliance.

Purpose of the Study:

• To introduce the Knowledge Retrieval Augmented Generation ENgine (KRAGEN) as a solution for complex problem-solving using LLMs.
• To mitigate LLM limitations such as hallucinations and improve the accuracy of generated insights.

Main Methods:

• KRAGEN combines knowledge graphs, Retrieval Augmented Generation (RAG), and advanced prompting techniques.
• Knowledge graphs are converted into a vector database for efficient fact retrieval via RAG.
• Employs graph-of-thoughts (GoT) prompting to decompose problems, retrieve relevant knowledge, and synthesize solutions.

Main Results:

• KRAGEN effectively limits hallucinations by grounding responses in retrieved factual knowledge.
• The system dynamically breaks down complex problems into manageable subproblems for systematic resolution.
• Provides a graph visualization for users to assess the logic and structure of the problem-solving process.

Conclusions:

• KRAGEN offers a robust framework for enhancing the reliability and accuracy of LLM-driven problem-solving.
• The integration of knowledge graphs and advanced RAG techniques represents a significant advancement in AI.
• KRAGEN is available as open-source, facilitating further research and application development.