Ask your own table in plain words

An export from a CRM or a till is a thousand rows with one question a week attached to it. Building a report for that takes a while, and opening the table and filtering by hand takes longer than asking.

What is inside

1. A question about the export — the entry.
2. Workspace storage — the node that says where files live.
3. Read the export reads the CSV by its path inside that storage.
4. Answer from the table — the model, given both the question and the rows.

A node reads the file rather than the model’s eyes: you can update the export without touching the scenario, because it reads whatever sits at that path today.

What to set up for yourself

• The file path on the read node: replace REPLACE_ME with something like sales-2026-08.csv.
• Delimiter and encoding. Exports from many systems arrive semicolon-separated and in a legacy encoding; both fields sit next to each other.
• The row limit (nrows_limit). It keeps both cost and accuracy in check: the whole table goes to the model as text. Five thousand rows is already expensive; beyond that, count with the df_* nodes and leave the model to phrase the answer.
• The prompt asks it to name the numbers the answer rests on. That is the only way to check it without opening the file.

What to connect

No connections; a file. Upload the export on the Files tab — that is the workspace’s own storage and it already exists. Updating the export means overwriting the same filename.