Databricks improves agent retrieval accuracy using structured chart extraction

Databricks details an approach that significantly improves agent retrieval and answer accuracy for chart-based questions within proprietary documents. The method leverages structured chart extraction capabilities from `ai_parse_document`, combined with lightweight text retrieval, outperforming large multimodal embedding models. This is crucial for enterprises, particularly in financial services, where agents often struggle to interpret data in figures and charts. The technique uses Databricks' AI functions to build a chart-aware retrieval pipeline.