FivetranChat: A homebrewed generative AI story

Fivetran has been exploring the potential of data integration to make generative AI practical since its emergence with ChatGPT. Retrieval-augmented generation (RAG), a core architecture supporting most current commercial implementations of generative AI, is a potent yet accessible way to turn proprietary data into useful, intelligent products. RAG requires two kinds of data integration: moving data from sources like applications and operational systems to data warehouses and data lakes, and turning data into embeddings that can be read by a vector database. Fivetran's internal chat tool, FivetranChat, is an example of how organizations can augment foundation models with proprietary data from text-rich systems such as CRMs and customer support applications for various use cases. The architecture involves extracting and loading data from every text-rich source closely involved with the product and operations, denormalizing it into text-rich tables using dbt transformations, and setting up a retrieval model with native AI tools like Snowflake Cortex and Databricks Mosaic AI to support generative AI. Overall, FivetranChat has been successful in providing accurate answers to internal questions about the product, operations, and company policies, reducing manual sifting through documents or asking for another person's time.