New Framework Cuts Enterprise Chatbot Query Time in Half by Automating Database Schema Linking

Conversational AI systems powering customer support are hitting a critical bottleneck: connecting customer questions to the right data sources when companies lack pre-built API access. A new approach addresses this friction point.

Researchers have introduced SAFAARI (Schema-Aware Framework for Accelerated Advertiser Response Intelligence), a multi-agent system designed to bridge the gap between natural language queries and SQL database retrieval. Rather than requiring engineers to manually map questions to database schemas, the framework automates this laborious translation step through coordinated specialized agents—one handling content, another managing metadata, and a third orchestrating the workflow.

How It Works#

The system relies on a novel evaluation metric called SEAL (Schema Evaluation and Accuracy in Language-to-SQL) that measures both accuracy and consistency across query responses. Testing five different feature configurations, SAFAARI achieved an 81.66% SEAL score—a 6.65% improvement over baseline approaches. The gains extended to individual components: datapoint accuracy improved by 5.51% and the precision of schema linking (correctly identifying which database tables to query) climbed 4.69%.

Real-World Impact#

Domain experts validated the framework through hands-on evaluation, confirming its ability to adapt across different support domains and data structures. The practical payoff: development time plummeted by a factor of eight compared to manual schema linking workflows, while accuracy remained high.

By automating what had been painstaking manual work, the approach lets customer support teams self-serve data queries rather than waiting for engineering intervention. It's particularly valuable for enterprises managing sprawling, complex databases where building custom APIs would otherwise consume significant engineering resources.