Data Retrieval in Transportation: Retrieval-Augmented Generation with Large Language Models

Abstract: In the era of big data, the rapid growth of data across various sectors has created an urgent need for advanced techniques to manage and optimize large datasets. Effective data retrieval methods are crucial for addressing complex queries, particularly in transportation. Retrieval-augmented generation (RAG) has emerged as a promising approach to enhance the capabilities of Large Language Models (LLMs) to retrieve information from proprietary or non-public data. However, how effective RAG might be in retrieving information from transportation related data remains unknown. This study investigates the application of RAG for accessing extensive transportation databases, focusing specifically on querying and retrieving Annual Average Daily Traffic (AADT) data for various road segments in Arizona across multiple years. The Arizona Department of Transportation (ADOT) stores AADT data for each year separately, and RAG was employed to identify the appropriate file for a given year and retrieve AADT values for specific road sections. Tests were conducted by providing prompts to assess RAG's ability to pinpoint the relevant year and road segments while delivering accurate AADT information for the specified year and segment. Various retrieval methods were evaluated, including Cohere Embeddings, Sentence Transformer MiniLM Embeddings, BM25 Sparse Retriever, and an ensemble of MiniLM and BM25. Overall, this research highlights the potential of RAG to significantly enhance data retrieval processes. Using a single prompt, RAG can effectively retrieve data from large transportation files, improving data accessibility and optimizing query responses. This advancement could lead to notable improvements in system performance across various applications. Continued development and refinement of these methods could unlock even greater potential for managing the increasing complexity of data within the transportation sector.