Skip to main content
The agent can write and run Python in a secure sandbox. It uses this to work with structured data. It can pull financials, prices, and holdings as tables, then filter, aggregate, and compute over them. This answers quantitative questions that search alone cannot. When you turn charting on, the same step can turn the results into a chart. The agent sends each chart back as a CHART message. This works on both the Research Agent (/v1/research-agent) and Workflows (/v1/workflow/execute).

​
Security

Each request runs in its own sandbox on a fresh process. No state is kept between requests. The code the model writes never sees your credentials. It cannot reach arbitrary external services. It can only read Bigdata’s own governed data APIs.

​
Usage

Code execution can run several analysis steps in one request. It can therefore use more content than a plain query. We monitor and optimize this with internal evaluations to keep it efficient. You can turn the tool off per request if you prefer.

​
Configuration

Configure code execution under the python_code_execution key of tools_configs:
  • enabled sets whether the agent may run Python. Set it to false to turn code execution off for a request.
  • chart_generation_enabled sets whether the agent may produce charts. Charts are off by default. Leave this false if you only want computed results in the answer text and no CHART messages.
For Workflows, you can store the same python_code_execution config on a template’s tools_configs. Every run of that template then uses it.

​
Charts

When chart_generation_enabled is true and the agent decides a chart helps, it sends a CHART message. Each CHART message includes a vega_lite_spec. This is a standard Vega-Lite specification. You can render it with any Vega-Lite renderer. Each message also includes start and end offsets. These point to a span of the answer text, so you can place the chart where the answer refers to it. This is the same offset model used by grounding citations.

​
Example

Take a request such as “Chart Tesla’s total quarterly revenue for the last six reported quarters.” The agent pulls the figures from Bigdata’s financials data, computes the series, and returns a CHART message. Rendering its vega_lite_spec produces this chart:
Bar chart of Tesla quarterly revenue
The chart is rendered from the vega_lite_spec in the CHART message.