Skip to content

CSV inspect

Inspect a delimited data file (CSV, TSV, semicolon or pipe separated) and get a structured report of its shape.

CSVTSVtableinspect

Upload

Files (0–1)
One delimited file, up to 8 MB, or paste a sample below.
Paste a delimited document instead of uploading a file. The same inspection applies; the source is never modified.

Options

Rows to sample 1–100000 Per-column type inference covers this many data rows.
Questions people ask
Will the tool modify my file?
No. csv-inspect only reads the document and writes one report. The source file is never re-serialised or returned in a transformed form, and no bytes of it reach the report beyond the header row names (or an explicit, off-by-default sample).
What does the report carry?
The inferred delimiter with a confidence grade (high, medium or low); header detection (auto, declared or undetect); the exact row count (data rows when a header is detected, everything otherwise; the counting rule is stated in the `notes` field); the modal column width and a count of outlier rows; per-column type guesses over a bounded sample; duplicate-header detection; a flag when the tool's sample ran out of rows before the file (`inference_partial`); and, if you asked for it, a bounded sample of data rows (off by default).
What does "per-column type guess" mean?
The tool walks the sampled data and, per column, checks whether all the values are integers, floats, booleans, text or a mix. The answer is one of int, number, bool, string, mixed or empty, with a confidence grade and, for mixed, the number of the data row where the first value broke the verdict. Cell contents never reach the report without the opt-in sample.
Are the heuristics authoritative?
No. Delimiter detection, header detection and type inference are heuristics, and the report labels them as such (`delimiter_confidence`, `header_detection` and the per-column confidence field). The exact counts, such as the width histogram, the outlier count and the row count, are stated as facts.
What happens with a file that is not a table at all?
A binary file becomes a successful inspection whose verdict is "not a text file" (delimiter: none, rows: 0, columns: 0). An empty file is reported as "empty". A single-column text file (no candidate delimiter is consistent) is reported as "unknown delimiter", with a confidence flag that tells a reader the tool is not pretending it found a table. In every case, the report is the answer and the job does not fail.
Does anything leave my server?
No. The tool runs in an isolated container with no network access. The `sample` in the report is bounded to ten rows of 120 characters per cell, and is off by default. Header row names (which can carry personal data, such as a user e-mail column) are always in the report, because a table report without its headers is not much of a report. If that is a problem for your data, strip the header row before uploading.