Reference / File types
What it reads, and how well.
- Formats accepted
- 41
- Fidelity tiers
- 3
- Refusals
- stated before upload
Tier 1 · Read directly
The file is the data.
Nothing is inferred and nothing is lost. Types, precision and structure come across as they were written.
23 formats
Tier 2 · Data extracted
Data inside something else.
The numbers live inside a layout built for a human reader. Structure is rebuilt, and every decision it makes is shown to you.
12 formats
Tier 3 · Read by eye, then checked
An image of a number.
Each figure is checked against the totals printed beside it. Anything the model is not confident about is flagged, never guessed.
6 formats
Spreadsheets
The common case. Every sheet is read, including the ones that are notes.
Every sheet read, including the ones that are notes rather than data.
Read the same as a modern workbook.
Values and formulas are read. Macros are never executed.
LibreOffice and OpenOffice workbooks, read like Excel.
Tables are pulled out of the Numbers package sheet by sheet.
Delimiter, quoting and encoding detected rather than assumed.
Read as delimited text.
Documents & decks
Reports written for people, with tables buried inside them.
Checked for a text layer first. A PDF written by software keeps its table structure; a scan is read as an image and told apart from the first kind.
Tables are lifted out of the document; the prose around them is kept as context.
Tables and the data behind native charts are read slide by slide.
Tables extracted, formatting discarded.
Database & app exports
Whatever your warehouse, API or admin panel hands over.
Nested objects are flattened into tables, keeping the path as the column name.
Streamed a record at a time, so an export of any size is fine.
Repeated elements become rows; attributes become columns.
Read columnwise, with the file's own types kept.
Every table read, with the foreign keys used to work out how they relate.
Read as SQLite when the file is one, and reported plainly when it is not.
Dashboards
The data behind a report someone else built: not a picture of it.
The tables inside the report are read out of its model, along with the measures that were defined on top of them, so your own definitions are used rather than re-invented.
A template carries the model and the measures without the data, and is read for both.
The packaged extract is read as data, and the workbook beside it supplies the calculated fields.
Calculated fields and the shape of each sheet are read.
Read straight through, table by table.
Bank & accounting
What a bank will actually give you, with the reference codes intact.
Transactions come with their own type codes and references, which makes matching far better than a CSV.
Read as OFX, which is what it is underneath.
Read as OFX, with the QuickBooks account mapping kept.
The SWIFT tags are parsed into transactions, balances and references.
Survey & research
Files that carry their own codebook, so the labels survive the trip.
Variable labels, value labels and missing-value codes are all read, so a 5 stays “Strongly agree” and a 99 stays missing.
Read the same as .sav.
Read the same as .sav.
Variable and value labels are kept, so the codebook does not have to be re-typed.
Read with its formats, including dates.
The interchange format SAS itself exports. Read like a .sas7bdat.
Screenshots & scans
A photograph of a printed sheet is a data source like any other.
Screenshots of a dashboard or a table are read, and every figure is checked against the totals in the same image.
A photograph of a printed report or a handwritten sheet is read into a table.
The same as .jpg.
Accepted as it comes off the phone. Safari usually hands it over as a JPEG.
Read like any other screenshot.
Multi-page scans are read page by page.
Notes & bundles
Context for the analysis, and a way to send a whole folder at once.
Read as notes, or as data when the contents turn out to be a table.
Markdown tables become tables; the rest becomes context for the analysis.
Unpacked, and each file inside is treated as though you had uploaded it yourself.
Every format above is accepted today. The ones marked as definitions-only are the honest exceptions: the file genuinely does not contain rows, so there is nothing to analyse, and that is better said here than after you have uploaded.
41 formats across 8 categories, 15 of which have a written guide.