A Retrieval-Augmented Natural Language Interface for Data Description and Meta-Analysis in the Pathogens-in-Foods...

Area of Science:

• Food Safety Science
• Computational Biology
• Data Science

Background:

• Food-safety occurrence databases are crucial for public health surveillance and risk assessment.
• Routine use is limited by the need for specialized database literacy and statistical programming skills.
• Existing databases often lack user-friendly interfaces for broad accessibility.

Purpose of the Study:

• To develop and evaluate a natural-language interface for querying food safety databases.
• To support grounded querying and reproducible evidence synthesis.
• To assess the performance of various language models in this interface.

Main Methods:

• Development of a retrieval-augmented natural-language interface with Open Chat and Guided Meta-Analysis modes.
• Utilized the curated and harmonized Pathogens-in-Foods (PIF) database.
• Evaluated four compact language models (Phi-4 Mini, DeepSeek-R1 Tool-Calling, Cogito, Qwen 3) and Gemini 2.5 Pro.

Main Results:

• All evaluated language models achieved 100% accuracy in tool selection and retrieval correctness.
• 100% argument extraction F1-score was achieved for relevant queries, demonstrating reliable database operation grounding.
• A case study on Toxoplasma in meat showed 100% numerical concordance and high visual informativeness, with Qwen 3 achieving the highest report quality index (93%).

Conclusions:

• Hybrid, evidence-grounded analytical interfaces can accelerate surveillance-oriented evidence synthesis in food protection.
• These systems, leveraging curated data and deterministic statistical backends, democratize access to critical food safety information.
• Performance variations were mainly due to language model interpretation quality, not tool execution failures.