데이터 추출

이 도구가 하는 일

This reads a set of documents and returns one row each, carrying the fields you asked for. Text is extracted in your browser and sent to the AI provider you choose with your own key; the files themselves never leave your device. The output is a JSON array or a CSV, with a _file column so every row says which document it came from.

It exists for the batch: twenty supplier invoices that have to become twenty rows in a ledger, a month of receipts before an expense claim, bank statements you want opening and closing balances from, or a folder of ID documents whose numbers need typing into a system.

작동 방식

1. Drop up to twenty PDFs onto this page, then choose your provider and paste your API key.
2. Pick a preset - invoice, receipt, ID document or bank statement - or choose your own fields and list the names you want, one per line.
3. Leave the missing-fields toggle on. It is what makes an absent value come back empty instead of invented.
4. Choose JSON or CSV, then press Extract. Each document is one request, and progress shows which one is running.
5. Download the file. Every row has the same columns, in the order you asked for them, whether or not each document had them.

Leaving missing fields empty is the most important setting on this page, and it is on by default. Asked for an invoice number, a language model will produce one whether or not the document has one, and a plausible invented invoice number is far more damaging than a blank cell - it is wrong in a way nobody notices until a payment goes to the wrong place. With the setting on, the instruction is explicit: where a value is not clearly present, return nothing, and never estimate. Turn it off only when you want values the document implies but does not state.

Every row comes back with exactly the keys you asked for, in the order you asked for them, present in the document or not. Missing values become null in JSON and an empty cell in CSV. That is what makes the CSV open cleanly in a spreadsheet: rows of differing shapes make columns drift, and a total in one row ends up under a date in the next. If a model's answer cannot be parsed at all, that row is filled with blanks rather than aborting a twenty-file batch at file eleven.

Temperature is fixed at zero, because extraction is not a creative task and the same invoice should give the same answer twice. Numbers are requested as bare numbers, without currency symbols or thousands separators, and dates as ISO 8601 strings, so the CSV sorts and sums without a cleaning pass. The CSV is quoted to RFC 4180, which stops a supplier name with a comma in it from splitting a column.

The text arrives in reading order with the layout flattened, and that is a real limit on what can be pulled out. A total in the bottom-right corner of an invoice reads fine; a multi-column table of line items arrives as a stream of cells with nothing marking which column each belonged to. It is why the presets ask for header-level fields - totals, dates, parties - rather than line items. For a table you genuinely need as a table, PDF to Excel reconstructs the grid instead.