A small note opens a larger question
Consider an illustrative invoice: components cost ₹12,000, freight is ₹800, and the total is ₹12,800. In the margin, a reviewer has written “Check freight against PO.” The referenced purchase order lists freight at ₹500. The discrepancy is ₹300, but the reason is not present on the invoice.
This is where structured extraction and retrieval solve different parts of the job. Assisto’s OCR engine reads the fields and handwritten note. Its retrieval engine then locates the relevant purchase order and supporting material. Neither an isolated transcription nor a search result is enough to answer the reviewer’s question.
Keep the row and its source together
The extraction schema includes the invoice reference, currency, line items, freight, total, and note. The purchase-order reference is used to retrieve the matching document rather than the most textually similar one. Values retain their page and row context so the reviewer can inspect where they came from.
A multi-page table can repeat headers or split a line item over a page boundary. A diagram may explain a delivery condition that is absent from the body text. Assisto’s document engine supports tables and illustrations; the workflow still needs to preserve those relationships when it assembles evidence.
A grounded answer stops at the evidence
A defensible response is: “The invoice charges ₹800 freight; the purchase order specifies ₹500, a difference of ₹300. No amendment was found in the supplied collection.” The answer links both source locations and names the missing evidence. It does not label the invoice fraudulent or automatically reject it.
Multi-hop reasoning connects the invoice to the purchase order and then to any amendment or delivery terms. Web research can help with external questions, but it cannot establish whether this buyer approved a private change. In this scenario, internal documentary evidence is decisive.
From explanation to a review queue
The proposed output contains extracted fields, a ₹300 discrepancy flag, and the source-linked explanation. An accounts-payable reviewer can accept a documented amendment or request clarification from the supplier. The business decision remains separate from the extraction.
Evaluation should include illegible notes, a wrong document version, missing amendments, and tables with repeated values. Measure field extraction and answer grounding separately: a correctly read ₹800 can still produce a wrong conclusion when compared with the wrong purchase order.
