How it works
- 1
Read the document structure
The extractor reads words and their positions, detects page rotation, and identifies the transaction table and its column headers.
- 2
Extract and normalise fields
Dates, amounts and currencies are parsed with locale-aware rules, while descriptions keep their text but lose layout artefacts.
- 3
Validate before you use the data
Running balances, duplicate rows and row completeness are checked. Anything uncertain is flagged and kept for review.
What gets extracted
| Field | Source in the statement |
|---|---|
| Transaction date | The posting date column or inline date |
| Value date | A separate value-date column or label |
| Description | The transaction narrative, including wrapped lines |
| Reference | Cheque, transfer or payment reference |
| Debit / Credit | Separate amount columns, DR/CR markers or sign |
| Balance | The running balance column |
| Currency | An ISO code, symbol or per-row currency |
How rows are reconstructed
PDF text arrives as individual words with coordinates. FlowParse builds geometric lines from those words, finds the table anchors, and decides which horizontal band belongs to one transaction. A description that wraps onto a second line is merged with its parent row instead of becoming a new transaction.
When a statement has no header row, the parser falls back to layout inference using date and amount positions. That fallback is conservative: if the structure is not clear, the affected lines are reported as unresolved rather than guessed.
Debits, credits and running balances
Some statements use separate debit and credit columns. Others use a single amount column with DR/CR markers, a minus sign or parentheses. FlowParse uses the document's headers and markers first, then the running balance to infer direction when the layout is ambiguous.
A running balance is also a validation signal. If the balance jumps in a way that the extracted amount does not explain, the row is flagged for review.
Validation checks
- Opening balance plus credits minus debits equals the stated closing balance.
- Each row has a date, a description and at least one amount.
- Consecutive running balances agree with the transactions between them.
- Potential duplicates are detected by date, description and amount.
- Ambiguous dates, currencies and amounts are flagged instead of silently chosen.
What you can do with the extracted ledger
Review and edit the rows in the workspace, then export to Excel for analysis or CSV for an import. Extracted statements remain available in your history, so you can revisit a month without re-uploading the PDF.
If you need the original document, FlowParse keeps the PDF in private storage and streams it back only to the owning account.
Extraction use cases
Financial analysis
Build a transaction dataset for cash-flow analysis, forecasting or category reporting.
Audit and investigation
Turn a stack of statements into a searchable, sortable ledger.
Expense review
Find recurring payments and unusual debits across several months.
Data migration
Move historical statement data into a new system that accepts CSV.
Questions
Does the extractor work on scanned PDFs?
Yes. Scanned pages are processed with OCR first, then the same extractor runs. OCR pages are marked and their rows are subject to confidence flags.
Can it extract from a statement with no table headers?
It attempts layout inference from date and amount positions. If the structure is too ambiguous, the affected lines are reported as unresolved rather than turned into invented rows.
Does it calculate a balance if the statement omits one?
No. The balance column is populated only when the statement shows a running balance. FlowParse does not fabricate a balance it cannot verify.
Can I edit extracted transactions?
Yes. The workspace editor supports updating a row, adding a manual row and deleting a row. Edits are saved through a security-invoker function under your account's permissions.
What export formats are available?
Excel (XLSX) and CSV. Both use the reviewed ledger, so what you corrected in the workspace is what you export.