Skip to main content
PrettyWhale.ai’s discovery engine reads a sample file you provide and uses it to infer field names, data types, nullability patterns, quality characteristics, etc.. The quality of that analysis, and therefore the quality of the generated pipeline, depends directly on how well your sample represents the real data. This page describes the supported file formats and how to prepare a sample that gives the model everything it needs.

​
Supported formats

PrettyWhale.ai currently supports the following sample file formats.
  • JSON
  • CSV
If your data is in a format not listed above (for example: Parquet, Avro, XML, or a fixed-width text file), contact support at contact@prettywhale.ai.The team can advise on format conversion options or confirm whether adjustment can be made to support your format.

​
Best practices for preparing sample files

A well-prepared sample produces sharper statistical analysis, more accurate type inference, and more comprehensive generated code. Follow these guidelines before uploading:
  • Include at least 100 to 200 records: A larger sample gives PrettyWhale.ai enough data to calculate meaningful rates, detect formatting inconsistencies, and produce representative quality statistics. A five-row sample will work, but the generated quality rules will be less precise.
The sample file can not exceed 50k max.
  • Include edge cases: Make sure your sample contains nulls, empty strings, unusual values, and boundary numbers (for example, zero-value amounts or maximum-length strings).
  • Use real production data (anonymized if necessary): Synthetic or hand-crafted samples often miss the quirks that exist in live data (mixed date formats, inconsistent casing, trailing whitespace). Anonymize personal data fields if needed, but keep the structure and value distributions intact.

​
Example files

Here is an example of a valid sample files if you want to test the solution: sample.zip