Inquiry on Document Intelligence Capability for Identifying Multiple Records in a Single PDF

PALAVALASAB
Tera Contributor

Hi Team,

I would like to understand whether Document Intelligence can identify and extract multiple records from a single PDF document, enabling the system to automatically process and close the corresponding records without manual intervention.

Specifically, I would like to know:

  • Can Document Intelligence detect and separate multiple records contained within a single PDF?
  • Can it match the extracted information with existing records and process them automatically?
  • How are unmatched, incomplete, or ambiguous records handled?
  • Is there a mechanism to flag such records for manual review while allowing successfully matched records to be processed and closed?

Our objective is to automate the handling of documents containing multiple records while ensuring that unmatched or uncertain records are routed for review and validation.

Please let me know if this functionality is supported and any prerequisites or limitations we should be aware of.

Thank you.

1 REPLY 1

kmohammed
Tera Guru

Short version: treat Document Intelligence as the extraction step only, and build the matching, closing, and review routing around it in Flow Designer. It will not, on its own, split one PDF into N records, match each one, and safely close them.

 

1. Separation: Document Intelligence extracts fields from a document into a task/record. If one PDF contains several logical records, design for that up front: either split the PDF into one document per record before extraction (where the source format allows a reliable split, for example by page or by a repeating section marker), or define the extraction so each repeated section becomes its own extraction output and then create one child record per output. Do not expect a single extraction pass to reliably invent record boundaries that are not marked in the document - test this with real samples, including the messiest ones, before promising full automation.

 

2. Matching and closing: matching is a separate step you build, not something extraction does for free. After extraction, look up the target record by a stable key from the document (a number, ID, or reference that is actually printed on it). Only auto-process when there is exactly one match and the required fields are present. Anything else - no match, multiple matches, missing key fields - must not auto-close.

 

3. Review routing: that exception path is the design, not an afterthought. Route unmatched, incomplete, or ambiguous items to a review queue or task, with the extracted values and the original document attached, so a person can correct and re-submit. Successfully matched items continue to processing/closure in the same flow, independently - one bad item in the PDF must not block the good ones.

 

4. Guardrails before going live: run a pilot on a representative sample and measure extraction accuracy and match rate by document type. Set a rule for what "good enough to auto-close" means (unique match + all required fields + a validation pass), and keep manual review as the default for everything that fails it. That is what makes the "automatically close the good ones, review the rest" objective safe in practice.