Agentic Multi-Database Question Answering

Schema linking, structured reasoning, and Text-to-SQL across heterogeneous databases.

Current work on an agentic data-analysis system that answers questions spanning multiple databases. The system decomposes a request, identifies relevant databases, tables, and attributes, produces structured intermediate decisions, and coordinates Text-to-SQL execution.

Using Qwen2.5-14B with tool calling and constrained outputs improved schema-matching performance by 20% in ongoing experiments. Modular validation and evaluation components make linking failures traceable and support reliable execution over heterogeneous schemas.

Stack: Python, Qwen2.5-14B, LangChain, LangGraph, SQL, structured outputs