Leveraging Retrieval-Augmented Generation to Accelerate Discoveries on Mealworm Larvae and Plastic Degradation

Area of Science:

• Environmental Science
• Artificial Intelligence
• Biotechnology

Background:

• Large language models (LLMs) show promise in scientific research but face trustworthiness concerns.
• Mealworm-mediated plastic degradation is an emerging field requiring efficient knowledge extraction.

Purpose of the Study:

• To evaluate Retrieval-Augmented Generation (RAG) techniques for improving LLM knowledge extraction in mealworm plastic degradation.
• To compare the performance of different RAG methods (GraphRAG, LightRAG, traditional RAG) with various LLMs.

Main Methods:

• Integrated 75 publications up to June 2024.
• Evaluated model performance using a curated dataset of 100 queries.
• Tested five LLM models (GPT-4o, GPT-5, Deepseek-V3.1, Qwen-plus, Llama-3.3) with different RAG approaches.

Main Results:

• LightRAG significantly improved LLM information extraction accuracy, exceeding 92% for quantitative data.
• LightRAG combined with Llama-3.3 provided the best balance of precision and coverage for open-ended queries.
• Empirical validation confirmed the accuracy of LightRAG + Llama for mealworm gut microbiome and plastic degradation patterns.

Conclusions:

• RAG, particularly LightRAG, offers a reliable method for advancing LLM applications in environmental science.
• The expandable nature of RAG facilitates timely knowledge base updates.
• Future research should address challenges like conflict handling in scientific AI.