How FactSet cut model training and deployment from days to hours with Databricks AI Runtime
FactSet cut the time required to set up, train, and deploy its Text-to-Formula models from days to hours by adopting Databricks AI Runtime. By leveraging serverless GPU compute directly alongside existing Lakehouse workflows, the team can now continuously retrain fine-tuned open models every two to three weeks to convert natural-language queries into FactSet Query Language formulas.
* FactSet’s Text-to-Formula (TTF) system converts natural-language requests from investment professionals into FactSet Query Language (FQL) formulas. After an initial RAG approach hit a quality ceiling on complex formulas, FactSet made fine-tuned open models a core part of the architecture. * The team retrains models every two to three weeks as new data becomes available, making infrastructure setup and deployment speed important parts of the development cycle. * With Databricks AI Runtime, FactSet can select serverless GPU compute for each workload and train alongside its existing Lakehouse workflows, reducing the time to set up, train, and deploy a model from days to hours.