Synthetic Data

What synthetic data is

Synthetic data is a dataset’s rows generated from a schema you define, instead of rows you upload or bring in yourself. You describe the columns you want, their names, types, and constraints, and Future AGI generates rows that match.

Define this schema for a customer-support dataset:

Column	Type	Constraints
customer_query	text	Value: a realistic customer support question
sentiment	text	Categorical values: positive, negative, neutral
priority	integer	Value: 1 (low) to 5 (urgent)

Generation produces rows like:

customer_query	sentiment	priority
I haven’t received my order and it’s been two weeks	negative	4
Can I change the shipping address on my recent order?	neutral	2
Your product is fantastic, just wanted to say thanks!	positive	1

Each column has its own Property editor for exactly this: Min Length and Max Length on most column types, Value set to Categorical for a list of allowed values, plus custom properties for anything else. Those column properties, not the column’s description, are where allowed values and ranges live.

To generate your first synthetic dataset hands-on, follow the Generate synthetic data quickstart.

The generation config

Every synthetic dataset saves what you defined as the dataset’s own generation config:

• Columns: the schema you defined, with each column’s name, type, and description
• Row count: how many rows to generate
• Description: what the dataset as a whole should contain
• Objective: how you plan to use the dataset, so generation can match that goal
• Pattern: an example or format you want the generated rows to follow
• An optional Knowledge Base link

That’s what lets you reopen a synthetic dataset later, change a column or the row count, and regenerate without rebuilding the schema from scratch. Here’s how the config, the job, and the dataset’s rows and state fit together:

flowchart TD
accTitle: How synthetic data generation fits together
accDescr: A synthetic dataset holds its own generation config, its rows, and a state that shows Generating, Regenerating, or Failed. A generation job reads the config, optionally grounded by a connected Knowledge Base, fills in the dataset's rows, and drives its state.
subgraph DATASET["Synthetic dataset"]
  CFG["Generation config"]
  ROWS["Rows"]
  STATE["State · Generating, Regenerating, or Failed"]
end
JOB["Generation job · runs in the background"]
KB["Knowledge Base · optional"]
CFG -->|read by| JOB
JOB -->|fills| ROWS
JOB -->|drives| STATE
KB -.->|grounds| JOB

When a Knowledge Base is connected, generation grounds rows in its content instead of relying on the schema alone.

Editing vs regenerating

Action	What happens
Edit the config and save	Adds the columns and rows you added, drops the columns you removed, drops rows if you lowered the row count, and leaves every other column’s data as it is
Regenerate	Rebuilds all rows and columns from the config (destructive, see the warning below)

The saved config is what makes either possible: you’re never redefining the schema by hand.

When to use synthetic data

Reach for synthetic data whenever real rows are unavailable, risky to use, or lopsided for what you’re testing:

• No real data yet: you’re building a new feature and don’t have production rows to test against
• Privacy limits: real data carries PII you can’t put in a test dataset
• Edge cases: you need scenarios that are rare in real traffic, like an angry customer or a multilingual query
• Scale: you need thousands of rows to stress-test a prompt or eval
• Skewed data: real data leans one way (mostly positive reviews) and you need a more balanced set

While it’s generating, and when it fails

Generation doesn’t happen instantly. Because it runs as a background job, a synthetic dataset sits in a Generating state (or Regenerating, if you kicked off a rerun) with a live progress bar while the job works, and a Configure Synthetic Data button that reopens the schema.

If the job errors out, the dataset shows a Failed state instead, with the same button to fix the configuration and try again. See Dataset FAQ & fixes for what to check first.

Keep exploring