Retrieval Challenges in Low-Resource Public Service Information: A Case Study on Food Pantry Access

A team of researchers including Touseef Hasan, Laila Cure, and Souvika Sarkar has published a study exposing the severe challenges AI faces when retrieving information from low-resource public service databases. Focusing on the critical domain of food pantry access, their work demonstrates how fragmented, inconsistently formatted, and outdated information creates a 'low-resource retrieval environment' that hinders timely access to essential services. To investigate this, the team developed a specialized AI system that scrapes publicly available pantry data and employs a Retrieval-Augmented Generation (RAG) pipeline, allowing users to ask natural language questions through a web interface.

In a pilot evaluation using realistic, community-sourced queries, the system revealed significant limitations in robustness, particularly when handling vague or 'underspecified' user questions and attempting to ground its answers in contradictory or inconsistent source data. This ongoing research, detailed in the arXiv preprint 'Retrieval Challenges in Low-Resource Public Service Information: A Case Study on Food Pantry Access,' moves beyond theoretical benchmarks to test AI in a socially urgent, real-world scenario. The findings underscore that for AI to be truly effective in improving access to public resources, future research must prioritize robustness and reliability in these messy, unstructured data environments, not just performance on clean datasets.