Table extractionthat reads the footnotes.
Complex tables are where document extraction usually breaks: multi-page, nested, and annotated with notes that change what the numbers mean. MiruIQ extracts them row-accurately and reads them the way a careful human would.
From table image to clean rows.
Three capabilities decide whether table extraction survives real documents.
Structure recovery
Rows, columns, merged cells and continuation pages are reconstructed into one logical table, not per-page fragments.
Interpretive notes
A note above the table saying amounts are in thousands, or that zeros are cut off on purpose, changes how every value is read.
Schema mapping
Extracted rows land in your schema's structure: typed amounts, dates and identifiers, ready for validation.
The table is not the hard part; its context is.
Any modern model can read a clean grid. Real tables come with legends, footnotes, currency headers and abbreviations that redefine the cells, and naive extraction silently produces wrong data by taking them literally.
MiruIQ treats those annotations as part of the table. Extraction happens against your schema with the surrounding text in view, and validation recalculates whatever must add up, so a misread table fails a check instead of reaching your systems. The uncertain cases park for human review with the table highlighted in the page viewer.
Table extraction questions
What teams with table-heavy documents ask most.
Yes, that is the headline capability: interpretive notes above or below a table (amounts in thousands, zeros cut off on purpose, currency stated once in the header) are applied to the values during extraction. The output contains the numbers as they are meant, not as they are printed.
Tables that continue across pages are reassembled into one logical table; nested structures extract into nested schema fields; and irregular layouts (grouped rows, subtotal lines, mixed row types) are handled by describing the target structure in the schema rather than hoping the grid is clean.
Validation does the proving: row amounts are summed against stated totals, running balances checked for continuity, and counts compared against declared counts. Tables that fail park for review, where the reviewer sees the extracted rows next to the highlighted source region. Correctness is checked, not assumed.
See it on your own documents.
The 14-day free trial covers 200 documents: enough to define your schemas, run real files and measure the results before anyone calls you.