Building Cost-Efficient Agentic RAG on Long-Text Documents in SQL Tables

A cost-efficient Agentic RAG architecture is proposed for querying large text documents stored in traditional SQL tables without schema changes. The system uses a ReAct agent to intelligently route user queries to specialized tools. One tool handles SQL-based computations and aggregations, while another performs semantic searches using a vector database. This hybrid approach allows the system to answer simple computational questions, complex semantic queries, and hybrid requests that combine structured filtering with content-based retrieval. The architecture is demonstrated using a news article dataset, showing how the agent selects the correct tool pathway for different query types.