Loan File Document Splitting: Rules, Code and Labels
Last reviewed: November 2026
Loan file document splitting is the process of breaking one scanned loan bundle into separate, correctly labelled documents such as identity proof, bank statement, salary slip and property papers. Done well, nobody in operations has to page through a 60-page scan to find one mandate form.
This playbook is for heads of loan operations and chief operating officers at banks, non-banking financial companies (NBFCs) and housing finance companies. It gives you a document class table, split rules you can configure, tested Python code, an accuracy formula and a worked bundle, and shows how Digiqt's DocuEye reads each document once it is labelled.
At a Glance
| Item | Detail |
|---|---|
| Input | One scanned portable document format (PDF) file per application, often 40–80 pages |
| Output | One file per document, named by application number and document type |
| Core steps | Classify pages, find boundaries, clean, write files, review low-confidence pages |
| Key measures | Boundary accuracy and label accuracy |
| Product | DocuEye extraction, behind a classification step Digiqt builds |
How Do You Split a Scanned Loan Bundle Into Labelled Documents?
You split a scanned loan bundle by giving every page a document class and starting a new file at each document boundary, where the class changes or a first-page cue appears. This document class table is the starting rule set.
| Document Class | Page Cues | Typical Pages | Output Label |
|---|---|---|---|
| Application form | Lender logo, "Application Form", photo box | 2–6 | app_form |
| PAN (Permanent Account Number) card | "Income Tax Department", 10-character number | 1 | kyc_pan |
| Aadhaar (masked) | Masked 12-digit number, Unique Identification Authority of India text | 1–2 | kyc_aadhaar |
| Salary slip | "Earnings", "Deductions", employer name, month | 1 per month | salary_slip |
| Bank statement | Bank logo, account number, date-amount-balance columns, "Page x of y" | 3–20 | bank_statement |
| Form 16 | "Form No. 16", TAN (tax deduction account number) of employer | 2–8 | form16 |
| Sale agreement | Stamp paper, "Agreement for Sale", schedule of property | 8–30 | property_agreement |
| Mandate form | NACH (National Automated Clearing House) mandate layout, bank details | 1 | nach_mandate |
| Cancelled cheque | Cheque image with "Cancelled" across it | 1 | cancelled_cheque |
Mandate forms then move to NACH mandate form processing, and Aadhaar pages to Aadhaar masking compliance.
Are your teams still paging through 60-page loan scans by hand?
Extract Split Loan Documents With DocuEye
Send us three anonymised bundles, or ask about loan file document splitting for your operations hub.
Which Rules Should a Document Splitting Configuration Hold?
A document splitting configuration should hold eight rules, set once per product and reviewed monthly; values shown are illustrative.
| Rule | Setting |
|---|---|
| Start a new document | Class changes, or a first-page cue appears |
| Continue a document | Same class and "Page x of y" continues |
| Blank page removal | Drop pages with almost no ink, keep a log |
| Page rotation | Detect orientation and rotate before classification |
| Confidence threshold | Below 0.85, send the page to a human review queue |
| Mandatory documents | Flag a bundle missing PAN, bank statement or mandate |
| Naming convention | {application number}{label}{sequence}.pdf |
| Original bundle | Keep unchanged, linked to every output file |
How Should the Human Review Queue Work?
The human review queue should show a reviewer page thumbnails with the proposed boundaries and labels, so a correction takes one click. Reviewers confirm or move a boundary, change a label, and send the bundle on.
Every correction is worth keeping, because it shows which classes the model confuses. Digiqt can feed those corrections back into the classifier it builds, and a daily target, such as clearing the queue before 4 pm, keeps credit from waiting.
What Is Loan File Document Splitting?
Loan file document splitting is the step that turns a single scan into a set of labelled documents. It sits between capture and storage, and it decides how easily every later team finds what it needs.
Loan file document splitting vs loan file scanning backlog: the backlog is about getting paper scanned at all, while splitting is about what happens to the scan. It also differs from loan document vaulting, which stores and secures documents after they are labelled.
Why Does Poor Document Splitting Slow Loan Operations?
Poor document splitting slows loan operations because every downstream team repeats the search. Credit, disbursement, audit and collections all open the same bundle looking for different pages.
Paragraph 47(6) of the Reserve Bank of India's (RBI) NBFC Know Your Customer Directions, 2025 asks NBFCs to maintain account information so it can be retrieved easily and quickly. That paragraph sits in Chapter VII on record management, and a bundle that hides the KYC (know your customer) pages on page 37 meets it badly.
Files that arrive through WhatsApp document collection or a paperless loan branch need the same splitting, because borrowers often send many documents in one PDF.
How Do You Code Document Splitting by Page Ranges in Python?
You split a PDF bundle by page ranges with the open-source pypdf library once the classifier has produced a plan. This code was tested with pypdf 6.19 on Python 3.13.
import os
from pypdf import PdfReader, PdfWriter
def split_bundle(src: str, plan: list, out_dir: str) -> None:
reader = PdfReader(src)
os.makedirs(out_dir, exist_ok=True)
for label, first, last in plan: # 1-based, inclusive
writer = PdfWriter()
writer.append(reader, pages=list(range(first - 1, last)))
writer.write(os.path.join(out_dir, f"{label}.pdf"))
plan = [
("APP1023_kyc_pan_01", 1, 1),
("APP1023_bank_statement_01", 2, 7),
("APP1023_salary_slip_01", 8, 10),
]
split_bundle("bundle.pdf", plan, "split_out")The hard part is the plan, not the cutting: page classification decides every range.
How Do You Measure Loan File Document Splitting Accuracy?
You measure loan file document splitting accuracy with two ratios checked on a weekly sample.
- Boundary accuracy = correct document boundaries ÷ true document boundaries × 100.
- Label accuracy = correctly labelled documents ÷ all documents × 100.
In an illustrative sample of 50 bundles holding 600 documents, 582 boundaries were correct and 588 documents carried the right label. Boundary accuracy is 582 ÷ 600 = 97.0%, and label accuracy is 588 ÷ 600 = 98.0%.
Set targets and a fixed response for each measure; the figures below are illustrative.
| Measure | Target | Action if Missed |
|---|---|---|
| Boundary accuracy | 97% or higher | Add page cues for the classes with most errors |
| Label accuracy | 98% or higher | Add rules for the pairs of classes most often confused |
| Pages sent to review | Under 10% | Check scan quality at the branches sending the most |
| Bundles missing a mandatory document | Under 5% | Return the checklist to those branches |
| Track errors by class, because a missed bank statement boundary costs more than a mislabelled photograph. Read it alongside our guide to optical character recognition (OCR) accuracy for Indian loan documents. |
What Goes Wrong Most Often in Document Splitting?
Five problems cause most document splitting errors, and each has a simple rule that prevents it.
- Interleaved pages: a salary slip scanned between two bank statement pages. Use "Page x of y" and account numbers to rejoin the statement.
- Two accounts in one statement scan: split on account number, not only on class.
- Duplex scans: blank backs of single-sided pages. Remove them by ink density and log the removal.
- Several documents on one page: front and back of an identity card photographed together. Tag the page with both labels.
- Rotated or upside-down pages: detect orientation before classification, or the page lands in review.
Keep the original bundle untouched, so any split can be rebuilt if an auditor questions it.
How Does DocuEye Read Documents After Splitting?
DocuEye reads documents after splitting in four steps: upload documents, select a template, artificial intelligence (AI) processing and structured output. DocuEye accepts PDFs, images and scans, and its "Advanced OCR + logic-based extraction engine" turns each one into editable, exportable data.
Page classification and splitting are a step Digiqt builds ahead of DocuEye's customizable extraction templates, so every labelled file reaches the right template without manual sorting. The step uses the labels from your document class table.
Once a file is labelled, DocuEye applies built-in validation rules, handwriting recognition and smart field detection across its 50+ document types. Your operations team still owns the human review queue and the final file check before disbursement.
What Does a 60-Page Home Loan Bundle Look Like After Splitting?
A 60-page home loan bundle becomes eleven labelled documents and two discarded blank pages in this illustrative worked example.
| Pages | Document | Output Label |
|---|---|---|
| 1–4 | Application form | app_form |
| 5 | PAN card | kyc_pan |
| 6–7 | Aadhaar (masked) | kyc_aadhaar |
| 8–10 | Salary slips, three months | salary_slip |
| 11–22 | Bank statement, six months | bank_statement |
| 23–28 | Form 16, two years | form16 |
| 29–44 | Agreement for sale | property_agreement |
| 45–54 | Previous sale deed | title_chain |
| 55–56 | Approved building plan | building_plan |
| 57 | NACH mandate form | nach_mandate |
| 58 | Cancelled cheque | cancelled_cheque |
| 59–60 | Blank separator pages | removed |
In this bundle, the agreement for sale and the previous sale deed look alike on stamp paper, so a reviewer should confirm the boundary at page 45 first.
If manual sorting takes 12 minutes per bundle and review of the split takes 3 minutes, a hub handling 400 bundles a day saves (12 − 3) × 400 ÷ 60 = 60 staff-hours daily; all figures are illustrative.
Each labelled file then goes to its own DocuEye template: salary slips feed income fields, the bank statement feeds balances, and the mandate form feeds account and limit details for the repayment set-up.
Want every scanned bundle split, labelled and read before credit sees it?
Plan DocuEye Extraction for Labelled Loan Files
Bring one product's bundles for a pilot, or visit Digiqt to scope a page classification step with our team.


