Analyze scattered data by asking questions in a visual workspace
You can combine data from various sources into a single view, explore charts, and branch out your analysis path as you follow your questions. Data Formulator is a tool that combines natural language questions with visualization tools to manage the analysis process on a single screen. It addresses the problem of losing track of the analysis flow due to complex data connections and long conversation histories.
Changes in the 0.8 Beta Version
The recently released 0.8 beta 1 version integrates data loading, questioning, result review, and branching into a single flow. It makes it easier to connect files, local folders, databases, and platforms like Databricks. Chart recommendations, themes, and styling tools have also been improved, enhancing visualization quality. You can try this version early using pip install --pre data_formulator==0.8.0b1 or uvx data_formulator@0.8.0b1.
Data Connection and Memory
Data is scattered across various locations such as files, databases, warehouses, and BI tools, which takes a lot of time to connect. Data Formulator provides a common connection method that allows agents to access diverse sources. It also maintains data memory to remember relationships between sources. This helps reduce the issue of agents providing answers when relationships are not clear.
Managing Analysis Paths with Data Threads
As you explore data, new questions constantly arise, and previous answers lead to comparisons or new directions. Long conversation histories make it difficult to understand your current position and how you got there. Data threads allow you to branch questions and compare different paths. They help you discover deeper insights using visualizations without losing context.
Visualization Engine and Installation
Charts are built on Flint, an open-source visualization language. Flint compiles concise chart specifications into refined visualizations. It supports more than 30 chart types, and fine-tuning is possible through a style improvement agent. There are various installation methods. You can run it immediately with uvx data_formulator, or install it with pip install data_formulator and then run python -m data_formulator. If you use Docker, start it with the docker compose up --build command. After running it, open the http://localhost:5567 address in your browser.
Pre-Usage Checks
This tool can start with data in various formats, including CSV, TSV, Excel, JSON, screenshots, text, and databases. The analyst discovers sources, clarifies requests, and proposes a loading plan for the user to review. The desktop download version is an automated preview build and is not code-signed or notarized. When bypassing operating system security warnings, you must verify that you downloaded it directly from this repository's workflow artifacts or releases. If you want to contribute, you must agree to Microsoft's Contributor License Agreement (CLA).