Data Formulator

Data Formulator is a Microsoft Research project for exploring data with visualizations powered by AI agents. Analysts start from CSV, Excel, JSON, screenshots, text, or databases and combine natural language with UI interactions to build charts. It is MIT-licensed, with a hosted demo at data-formulator.ai and local installation options. The latest release listed is 0.8 beta 1.

What It Is

Data Formulator is an AI-assisted data analysis and visualization workspace. Users load data, ask questions in natural language, and review the results as explanations, tables, and editable charts. The README says charts are built on Flint, an open-source visualization language that compiles compact chart specs into visualizations.

How the Workflow Works

Data can come from uploaded files (CSV, TSV, Excel, JSON), screenshots or text, databases and data platforms, or from an analyst agent that discovers sources, clarifies the request, and proposes a loading plan for review. Questions are asked in the Data Thread, which keeps one history of the reasoning. Users can refine a result directly, branch from any earlier step to explore alternatives, or delegate the next investigation to the analyst. Finished analyses can be composed into a report to share.

Data Connectors and Models

Data connectors give agents a common way to reach different sources, and the system keeps a data memory of the relationships between them. Earlier releases added loaders for MySQL, PostgreSQL, MSSQL, Azure Data Explorer, S3, and Azure Blob, DuckDB support for large data, and Databricks is mentioned for 0.8. Model support covers OpenAI, Azure, Ollama, and Anthropic via LiteLLM.

Setup Path

The hosted site is a demo with example datasets and uploads. Local use unlocks database connections, local folders, persisted sessions, custom models, and user management. It can be run with uvx, pip, or Docker, and opens in the browser at localhost:5567. Preview desktop builds for Windows and macOS are available through CI artifacts and releases; the README notes they are not code-signed or notarized.

Current Status

The 0.8 beta 1 release (0.8.0b1) unifies loading data, asking questions, reviewing results, and branching in the Data Thread. It adds more data sources and more Flint-powered charts, recommendations, themes, and styling tools. The README names 0.7 as the latest stable release.