Document details

Scientific databases like the Pathogens in Foods (PIF) Database hold valuable public health data but are often inaccessible to experts lacking programming skills. This research addresses this gap by developing and evaluating a novel Visual Natural Language Interface (V-NLI) for the PIF database. The resulting PIF Intelligent Agent empowers users to perform complex queries, conduct meta-analyses, and generate dynamic reports using natural language. The agent uses a hybrid, dual-mode architecture separating language interpretation from statistical computation. An "Open Chat Mode" offers a flexible exploratory interface via a tool-calling Small Language Model (SLM) with Retrieval-Augmented Generation (RAG). A "Guided Meta-Analysis Mode" provides a structured workflow for generating reproducible scientific reports through a dedicated Rserver backend. A comprehensive evaluation benchmarked five SLMs: Phi-4 Mini (3.8B), MFDoom/deepseek-r1-tool-calling (14B), Cogito (14B), Qwen 3 (8B), and Gemini 2.5 Pro. While all models achieved flawless functional accuracy, their effectiveness was determined by interpretive quality. The ability to generate concise, factually coherent text was the key differentiator, with smaller, instruction-tuned models showing performance comparable or superior in conciseness to larger models. The end-to-end system proved highly reliable, validating the architecture and establishing interpretive fidelity as a critical benchmark for domain-specific agents.