---
title: "Statement import"
description: "Upload a statement as PDF, image or CSV, match its rows against your ledger, and see what is missing, duplicated or booked wrong."
sidebar:
  icon: file-text
---

The AI reads the rows; the app then compares them against your transactions and shows what is missing, duplicated or booked wrong.
## Which file types can I upload? [#which-file-types-can-i-upload]

- **PDF** — the digital statement from your bank. The text layer is extracted locally and only the text is sent to your AI endpoint.
- **Images** — PNG, JPEG, WebP or GIF, e.g. a **photo or screenshot** of a paper statement or a banking app. The image is sent to your AI endpoint as an image, so the connection you use for *Statement extraction* must support vision.

A scanned PDF without a text layer cannot be read — take a photo/screenshot of it instead and upload that as an image.

## How do I import one? [#how-do-i-import-one]

1. Open **Statements**, pick the **account** the statement belongs to.
2. Choose the file (PDF or image).
3. Set the **date tolerance** (default 3 days) — booking dates in the app and at the bank rarely match exactly.
4. Enable **invert amounts** if the statement shows expenses as positive numbers (common on credit-card statements).
5. Start the import. The AI extracts the rows, then a deterministic matcher compares them with your transactions.

## How does matching work? [#how-does-matching-work]

Matching is done in code, not by the AI:
- **Amount must agree to the cent.** Split transactions are summed per split group first.
- **Date** must be inside your tolerance window.
- **Description similarity** only ranks candidates and decides *exact* vs *probable* — it never creates a match on its own.
- Each app transaction can be consumed by **at most one** statement line, so repeated identical amounts are never double-matched.

## What do the result groups mean? [#what-do-the-result-groups-mean]

- **Missing** — on the statement, but not in the app. Create the transaction with one click (the Add form is prefilled).
- **Probable** — a likely match; confirm or reset it.
- **Matched** — exact matches, nothing to do.
- **Ignored** — rows you marked irrelevant (fees you don't track, carry-forwards).
- **Not on the statement** — transactions in the app inside the statement period that the statement does not contain: typically a duplicate, a wrong date, or a booking on the wrong account.

## What is sent to my AI provider? [#what-is-sent-to-my-ai-provider]

It depends on the file type. **CSV and TSV are read entirely locally — nothing reaches a provider at all.** For a PDF, the embedded text layer is extracted and sent in chunks; for an image, the picture itself.

A second pass then runs automatically to categorise lines that matching could not tie to a transaction. That pass also sends a **snapshot of your ledger** — accounts, categories and your recent transactions — because without that context it cannot guess where a line belongs. Matching itself runs on the server with no AI call.

Choose connections under **Settings → AI Assistant**, separately for *Statement extraction* and *Line classification*; the usual fallback to the next enabled connection applies.