Learn how we built Reffy, a full-stack agentic app on Databricks for finding customer references, and how you can use the same architecture to solve knowledge discovery at your organization.

databricks

Databricks built Reffy, an internal agentic app that consolidates over 2,400 customer references from YouTube, LinkedIn, Medium, internal slides, and the company website into a searchable, AI-powered tool. The architecture uses a medallion ETL pipeline (Bronze/Silver/Gold Delta Lake tables in Unity Catalog), a 31-point LLM-based quality scoring system via AI Functions with Gemini 2.5, Databricks Vector Search with hybrid keyword and semantic retrieval, a DSPy-based tool-calling agent deployed to Model Serving, and a React/FastAPI frontend hosted on Databricks Apps. Lakebase persists conversation history and logs, while Lakeflow Jobs power both ETL and monitoring dashboards. In its first two months, over 1,800 sales and marketing employees ran 7,500+ queries. The team plans to expose Reffy via an API and MCP server for broader workflow integration.

From Tribal Knowledge to Instant Answers: Building Reffy on Databricks

The challenge of democratizing tribal knowledge

Reffy: A full-stack solution on Databricks