Engineering drawing data extraction

OCR returns words. On a drawing the difficulty is not reading the characters but deciding which string is the drawing number, which is the revision, and which balloon number belongs to which part. AxioExtract asks the model for a defined schema per drawing type, which is why the output columns are identical on every sheet of a set and can be appended into one workbook.

Complex spatial sheets — drawings, schematics, isometrics and P&IDs — are routed to a frontier vision model, while text-based documents use a lighter model. CAD and BIM files skip vision models entirely and are parsed programmatically instead; see DXF to Excel and IFC to Excel.

Every extraction opens with the sheet beside the results grid. Selecting a row highlights where that value was read on the page, low-confidence readings are flagged, and any region can be boxed and re-read. Corrections are made in the grid, and the export reflects the reviewed values — not the raw model output.

A single issued sheet returns the title block first — drawing number, title, revision, scale, projection, material, finish and the drawn/checked/approved names and dates. Then the revision history as its own table, the general notes in printed order, and the dimensional and GD&T callouts with their tolerances. What was a picture becomes a register you can sort, filter and compare against the next revision.

Every value stays tied to the page: selecting a row highlights where it was read, flagged low-confidence readings are obvious, and any region can be boxed and re-read on its own rather than re-running the whole sheet.

Drawings, schematics, P&IDs and isometrics are spatial: the meaning of a value depends on what it points at. Those sheets are routed to a frontier vision model, which is why a drawing is 4 credits per page and a P&ID, schematic or isometric is 8 credits per page, while a plain office table is 1 credit per page. Where a document carries its text as data — DXF and IFC — no vision model is used at all.