Advisory, not automatic
Schema inference profiles accepted records and reports, per column, which types every non-empty value fits, the closest alternative and the first values that break it. It never converts. ZIP codes like “02134” look like integers, but converting them loses the zero; account IDs with 20 digits exceed what a JavaScript number can hold exactly. Both cases get a note.
Validation rules are not parse errors
A record can be perfectly valid CSV and still break your contract: a missing required value, a non-integer quantity, an amount out of range. The rule checker reports those as schema findings, in their own list, so “the file parses” and “the data meets our rules” stay separate questions.
Recognised types
| Type | Accepted text |
|---|---|
| integer | Optional sign and digits |
| decimal | Integer, decimal point number or exponent form (dot decimal separator) |
| boolean | true/false, yes/no in common capitalisations |
| date | ISO 8601 calendar date YYYY-MM-DD |
| date-time | ISO 8601 date and time with optional seconds and offset |
Questions
Does it understand decimal commas or local dates?
Not as numbers or dates. Values such as “0,42” or “03/04/2026” are reported as text, because their meaning depends on locale. You can still validate them with length rules.
Which records are profiled?
Accepted records: OK, with warnings and repaired. Rejected records are excluded so malformed rows do not skew the suggestions.