Document parsing and field extraction should be designed separately
Z.ai's GLM-OCR model card introduces it as a multi-modal OCR model for complex document understanding, with a parameter size of about 0.9B. The officially recommended SDK is used for document parsing, combined with layout analysis to form structured output; the information extraction task requires direct model inference according to the specified JSON structure.
Document parsing focuses on the original content and its layout, and field extraction focuses on the information required by the business. The two can be connected back and forth, but the example of parsing the page cannot be directly regarded as being adapted to any business form, nor should the general chat model be used to change the task prompts at will.
First define a table that is actually used (ChinaOPC analysis)
For example, when sorting out purchasing materials provided by suppliers, you can first agree on the product name, model, quantity, unit, data date and source page. Field definitions should describe the allowed types, how to represent missing values, and which values must be confirmed by humans; as long as the field names are the same, there is no guarantee that different data will use the same caliber.
When "Quantity per box" and "Number of purchased boxes" appear at the same time, they should not be put into the same quantity column. Currency, tax-inclusive description, unit of measurement and applicable batches should also be recorded separately. Before the model outputs a number, the system needs to know what the number actually describes.
Format verification is not a substitute for fact checking
JSON can be read by the program, which only means that the structure can be parsed. Also check the corresponding position of the field and the original text, the decimal point of the number, symbols and blank processing. Each key field can be associated with the page number and original image area, so reviewers can directly view the basis and reduce the time of searching page by page.
Values not found should be left empty or pending confirmation and cannot be filled in with common values, old orders, or contextual guesses. When an abnormal unit, duplicate model or conflicting information occurs, it will enter the waiting queue and be written to the formal business table after confirmation.
Design the import into a reviewable process
Each batch processing saves the original file list, extracted version, review status and final import record. When uploading the same file repeatedly, identify whether it has been processed first to avoid generating duplicate entries. Manual correction should leave the values before and after correction to facilitate locating whether the error comes from the source data or the identification process.
A small team can first deliver a fixed form type organization service and then expand to other layouts. The cost is included in manual proofreading and exception handling; this article does not measure recognition accuracy, nor does it describe OCR output as financial review or formal business approval.
Further reading
Understand the full page parsing pipeline; Understand Pangu industry data service. For more introduction, please see AI Model Column.