Your Data Questions Deserve Branches, Not a Never-Ending Chat Log
You've been there: you ask an AI agent about your data, get an answer, ask a follow-up, then another, and twenty messages later you're scrolling through a wall of chat history trying to remember which thread of reasoning actually led somewhere useful. Data Formulator is a Microsoft open-source project that takes a different approach, giving you a visual workspace where your data questions branch into separate threads instead of stacking up in one long conversation.
What It Does
Data Formulator is an AI-powered data visualization tool. You load data, ask questions, and get charts back, with AI agents doing the heavy lifting of turning your questions into visualizations. The project frames the problem it solves in two parts: data lives everywhere (files, databases, warehouses, BI tools), and questions evolve as you explore, which makes a single linear chat history a poor fit for real analysis.
Its answer is a unified visual workspace built around two pieces. First, data connectors give agents a common way to connect to different data sources and maintain a "data memory" that remembers the relationships between them. Second, Data Threads let you branch into different questions, compare paths, and dig into insights with visualizations without losing context. The charts themselves are built on Flint, a separate open-source visualization language from Microsoft that compiles compact chart specs into finished charts.
It's available both as a hosted online demo and as a local install via pip or uvx, and it's MIT licensed.
Why It's Cool
-
The threading model is the real idea here. Most AI data tools assume a chat interface is the right metaphor. Data Formulator argues that exploration is tree-shaped, not linear—each answer opens follow-up questions, and you want to see those paths side by side. Branching instead of appending is a small conceptual shift with a big effect on whether you can retrace your reasoning.
-
Data memory addresses a real agent failure mode. The README points out that agents often start answering before the relationships between data sources are clear. Keeping a memory of those relationships means the agent isn't guessing at how your tables connect every time you ask something new.
-
The connector story is practical, not theoretical. Files, local folders, databases, and platforms like Databricks are all in scope for the current beta. If your data doesn't live in one tidy CSV, that matters.
-
Charts are handled by a dedicated engine. Rather than rolling its own rendering, the project leans on Flint for visualizations, which the README describes as compiling compact chart specs into polished output. Version 0.7 mentions 30+ chart types refined by a style-refinement agent, and the 0.8 beta adds more Flint-powered charts, recommendations, themes, and styling tools.
-
It ships like a real tool. Sessions and workspaces persist, the UI supports English and Chinese, and you can share results as reports. Those aren't glamorous features, but they're the ones that decide whether something gets used past the first afternoon.
How to Try It
The quickest route is the online demo at data-formulator.ai. If you'd rather run it locally, you have two options.
Install the latest stable release (0.7):
pip install data_formulator
Or run it instantly without a persistent install:
uvx data_formulator
If you want to preview the 0.8 beta, which introduces the unified flow (load data, ask questions, review results, and branch in the Data Thread), install it explicitly:
pip install --pre data_formulator==0.8.0b1
or:
There's also a YouTube walkthrough linked from the repo, a changelog for release details, and a Discord if you want to ask questions. The source and full instructions live at github.com/microsoft/data-formulator.
Final Thoughts
Data Formulator is aimed at people doing exploratory analysis who've found chat-based data tools frustrating once the conversation gets long—analysts, data scientists, anyone who asks a lot of follow-up questions of the same dataset. The threading and data-memory ideas are the parts worth paying attention to, and the fact that the project is actively iterating (0.8 is in beta, with a broader set of data sources) suggests the team is still working out the design rather than coasting. If your analysis keeps turning into a tree of "what if I look at it this way instead," this is built for exactly that.