Skip to main content

How a file is read

Reading a spreadsheet and reading a contract are not the same job, so SnoutData does not decide for you. Right after you add a file, you choose how it gets read, and each option tells you where your data goes before you pick it.

That is a fact about the route, not a reassurance: the option you choose determines which machines see the content.

The four options​

They are listed in escalation order. Start at the top and move down only if you need to.

1. The parsers, on this computer​

The default. Structured files (CSV, TSV, JSON, NDJSON, logs) are read locally by code.

Where the data goes: nowhere. It stays on this computer. No model is involved at any point. This is the fastest option and it costs nothing.

2. A model on this computer​

For documents the parsers cannot handle, such as a PDF of prose. A local model reads them on your own hardware.

Where the data goes: nowhere. It stays on this computer. The model is downloaded on demand the first time you choose this option.

3. Our AI​

SnoutData's managed gateway, metered against your plan. Use this when the local model is not good enough for a document and you would rather not run one yourself.

Where the data goes: the content reaches a model through our gateway, so that it can be read.

4. A model of your own​

Your own API key and your own provider.

Where the data goes: straight from this computer to that provider. It does not pass through SnoutData's servers, and the billing relationship is between you and them.

A file is not handed over whole​

When a model does read something, it is not given the entire file.

A table has a regular shape, so SnoutData samples it: the header, the first rows and the last rows. The model is asked how to read this table, and then the parser reads every row that way. A one-gigabyte export costs a single call, and only a few dozen rows are ever exposed.

A document has no such regularity, so it is read in windows, in as many passes as it takes. There is no length at which a file is refused for being too long.

When a read fails​

A failure is reported, and the flow does not quietly fall back to a different reader. Two different things are kept apart:

  • The reader could not run at all (for example, the local model is not installed). The run stops.
  • The reader could not read this particular file. That file is named and skipped, and the run continues with the others.