Document processing
Table extraction
Define columns, or just describe what you need; the AI reads the document, extracts, and verifies every cell verbatim.
Table extraction
What happens to your table
Every cell that comes out of this tool is checked against the text layer of the page it came from. The model reads the table and writes values; the verifier then asks, for each value, does this exact text appear in the document? A cell it can confirm is delivered as a normal value. A cell it cannot confirm is marked, and the delivery as a whole is labelled as carrying warnings rather than reported as a clean success.
This check exists because a model asked to transcribe a table will sometimes improve it. In our own testing it turned a cell reading 循环贷款 into 循环贷款等 — one character added, invisible when skimming the output, and wrong. Schema validation cannot catch that, because the column is still the column and the type is still the type. Only comparing against the source can.
You can define the output columns yourself, or describe what you need in a sentence and let the tool infer the structure. Describe mode returns a free structure preview first, so you can see whether it understood the table before spending credits on a delivery. Each detected table becomes its own sheet in the workbook.
Frequently asked questions
- What happens when a value in the scan is blurred or unreadable?
- It is marked as unconfirmed rather than guessed silently. We do not ask the model to invent a plausible value and deliver it as fact, because a plausible wrong number in a spreadsheet is worse than a visible gap. The delivery is flagged so you know to check those specific cells.
- How do I know which cells were not verified?
- The result carries warnings that name the affected table, and any cell the document could not confirm is reported individually. A delivery containing unconfirmed cells is marked succeeded with warnings, never as a clean success, so the outcome itself tells you to look closer.
- Does it keep the row and column structure of the original table?
- Yes. The layout is read from the source geometry, so merged cells, row spans and the reading order are carried through rather than flattened into a single column. Each table becomes a separate sheet, so a document with several tables does not collapse them into one.
- How large a document can I upload?
- Up to 50 MB per source file. Tables are read per page, so a long report costs the same per table whether it is on page 2 or page 200. Splitting a huge scan into smaller files is usually faster than uploading one enormous one.
- What is the difference between this and an OCR program?
- OCR gives you characters in reading order and leaves you to decide whether they are right. This gives you a table structure plus a per-cell check against the source, and it tells you which cells it could not confirm. The output is a workbook you can audit, not a wall of text you have to re-read.
- What format do I get back?
- An Excel workbook with one sheet per detected table, plus a JSON manifest that records the source page count, the source checksum and any warnings. The manifest is what makes a delivery explain itself later, without needing to trust memory about which run produced which file.
What this tool does not do
It does not repair a scan. If a value is genuinely illegible in the source, no amount of processing recovers it; the tool will tell you the cell is unconfirmed rather than produce a number.
It does not create formulas. Values visible in the scan are transcribed; calculations the original performed are not reconstructed.
A confirmed cell means the text appears in the document text layer, not that the table was interpreted the way you intended. If you defined a column that does not exist, the tool cannot invent it, and the delivery will show that.