Synthetic Data Generation

1. Go to app.futureagi.com, then Dataset, then Add Dataset
2. Select Create Synthetic Data

You should see the Create Synthetic Data wizard open, prompting for a dataset name and description.

Select knowledge base (optional): select a Knowledge Base to ground the generated rows in your own documents instead of the schema alone. Leave it empty to generate without domain grounding. Give the generated rows a skim before you rely on them; column properties aren’t hard validation on the result.

To set up a KB first, see the Knowledge Base cookbook. You can also start directly from the KB detail view: click Create Synthetic data in the action bar, and the wizard opens with your KB pre-selected.

Click Next.

You should see the wizard advance to the column configuration step.

Add three columns using the Add columns button.

Column 1: question

• Column Type: Text
• Properties: Min Length = 20, Max Length = 200

Column 2: answer

• Column Type: Text
• Properties: Min Length = 50, Max Length = 500

Column 3: category

• Column Type: Text
• Properties: set Value to Categorical with:
  • shipping: 25%
  • billing: 25%
  • returns: 25%
  • account: 25%

Click Next.

You should see the wizard advance to the column description step, with an input for each column you added.

Write a description for each column. Use {{column_name}} to reference other columns: this creates dependencies so generated values are contextually related.

Column 1: question

A realistic customer support question about {{category}} issues.
Phrased as a real customer would type it in a chat widget.

Column 2: answer

A professional support response to {{question}} about {{category}}.
Directly addresses the concern with a clear next step.

Column 3: category

The support category this Q&A pair belongs to.

You should see all three description fields filled in, with the Create Dataset button now active.

Click Create Dataset. Generation runs in the background and the platform redirects you to the new dataset, which shows a Generating state until the rows land.

You should see the support-qa-synthetic dataset move from Generating to 20 rows across the question, answer, and category columns.

• Sort or filter rows to inspect quality
• To re-generate or modify: click Configure Synthetic Data in the dataset toolbar. The Synthetic Data Details drawer opens

• Re-Generate same Configuration: rebuild every row from the same settings (destructive)
• Edit Configuration: modify the schema, then choose:
  • Replace the current dataset: overwrite with new rows
  • Create as new dataset: keep the original, generate a separate dataset
  • Add it to existing dataset: append new rows

You should see the dataset either regenerate in place or a new dataset appear, depending on the option you chose.

1. Click Evaluate in the dataset toolbar
2. Add Evaluations then select completeness
3. Map keys: output to answer, input to question
4. Click Add & Run

You should see a completeness score column appear next to the generated rows. A row like this scores low (illustrative example):

The answer column description only says “Directly addresses the concern with a clear next step,” but nothing in the Pattern or description forces a concrete action (a link, a timeframe, a next step). To fix it, go back to Step 6, open Edit Configuration, and tighten the answer description to something like “State the specific next step the customer should take (e.g., the exact page to visit, or that a refund will process within N business days).” Regenerate and re-run completeness: a typical before/after on a tightened description moves scores from the 0.3-0.5 range up to 0.8+ (illustrative; your numbers will vary by run). Filter out any rows still scoring low before using the dataset for fine-tuning.