Files and bulk import

The assistant reads files as well as receipt photos, in the web Assistant panel and in the Telegram bot. A file is either one receipt or a list of transactions (a bank statement, a spreadsheet, an export from another app). A list is shown as one Confirm card, and Confirm saves every row at once. You can send up to 5 files together. They become one import, and a transaction that appears in two files is imported once.

Supported files

Format	What happens	Limit
PDF with text (an e-receipt, a bank statement)	The text is read. One purchase becomes a receipt card with the PDF attached. A statement becomes a bulk import.	8 MB, first 20 pages
Scanned PDF (just an image)	The first page is read like a receipt photo.	8 MB
Word (`.docx`)	The text is read, like a PDF. The file itself isn’t attached.	5 MB
CSV (`.csv`, `.tsv`, `.txt`)	Rows are read directly, without AI.	5 MB, 1000 rows
Excel (`.xlsx`), OpenDocument (`.ods`)	The first sheet is read, like a CSV.	5 MB, 1000 rows
OFX / QFX (`.ofx`, `.qfx`)	The bank export format (old SGML and new XML both work). Read without AI; the bank’s transaction IDs prevent doubles.	5 MB, 1000 rows
QIF (`.qif`)	Quicken / old money-app export. Read without AI; the category becomes the tag (`Food:Groceries` → Groceries).	5 MB, 1000 rows
JSON (`.json`)	A FundsFlow export (Settings → Download JSON) is read as is, tags included. Any other list of objects with a date and an amount is read like a CSV.	5 MB, 1000 rows
HTML (`.html`, `.htm`)	A saved statement page. The table with dates and amounts is read like a CSV; a page without such a table is read by the AI, like a PDF.	5 MB
Screenshots of a bank app’s transaction list	All screenshots are read together; transactions that show on two overlapping screenshots are kept once. See below.	up to 10 in Telegram, 5 on the web

Old or macro formats are refused with a hint: save .xls, .xlsm, or .xlsb as .xlsx or .csv, and .doc, .docm, or .rtf as .docx or PDF. Macros are never run; files with macros are refused. A .txt file that is really OFX, QIF, JSON, or HTML is recognised by its contents.

On the web, attach files with the 📎 button or drag them onto the panel. Each file shows as a chip with its name and size above the message box; remove one with ✕. Up to 5 files per message, 20 MB together, and not together with photos. In Telegram, send them as files (documents). Up to 5 files sent together (one album) become one import; files sent separately are read one by one.

A PDF receipt with an amount in the text (-350 groceries) is saved right away with the PDF attached, as before, and uses no AI.

Say what to change

What you write with the files is applied to the import, and the card lists it under 📝 From your message. For example:

only expenses, only income, only USD
skip the first line, skip Rent and transfers between my cards, only October
expenses are positive or flip the signs
dates are month first (or day first) when a file uses dates like 03/04/2026
tag everything Travel, put Uber and Bolt under Transport
create the tags if needed, without tags, Cafe → Food, don’t suggest tags

Ukrainian and Russian work too (лише витрати, створи відповідні теги при потребі, спочатку місяць). Common phrases like these are understood without AI. Anything else uses one AI assist, or none if the import already used one (a PDF statement or screenshots). A word or two (October, Monobank card) is treated as a label, not an instruction, and costs nothing. If something can’t be applied, the card says so.

Tags that don’t exist yet

Tags in the file (a Tags or Category column, QIF categories, a FundsFlow export) are matched to your tags by name, ignoring case. When some don’t exist, the assistant asks first:

🏷 These tags don’t exist yet: Cafe (12 rows), Taxi (3 rows). Create them, import without them, or map them to your tags (e.g. Cafe → Food)?

Press Create them or Import without tags (buttons on the web and in Telegram), or Cancel.
Or just write the answer: create Cafe, Taxi → Transport, Cafe → Food, the rest without tags. Other edits work here too (and skip Rent).
Then the Confirm card appears, with 🆕 Will create tags: … if any. Nothing is created until you press Confirm; new tags get the 🏷 emoji, which you can change later.
If you said create the tags if needed with the files, the question is skipped. After three replies that don’t settle it, the remaining unknown tags are left out.

Tags the AI only guessed (from a PDF or screenshots) are never created on their own: they’re kept if you have them. Rows that have no tags get suggestions from your history: when you’ve saved the same note before (Silpo), the tags you used most for it are added. These show with ~ in the preview. Turn this off with don’t suggest tags.

Screenshots of a bank app

Send screenshots of the transaction list from your bank app:

On the web, attach up to 5 screenshots and write statement (or transactions, виписка, список).
In Telegram, send up to 10 as one album. Write statement with them, or press 📋 Transaction list when the bot asks One transaction / Separate / Transaction list.
A single screenshot that shows a list (not a receipt) is recognised on its own.

All screenshots are read in one AI assist. When screenshots overlap, the rows they share are counted once. Rows without a visible date are skipped and listed. The screenshots themselves aren’t saved.

Bulk import preview

The Confirm card shows:

how many transactions, their date range, and which files they came from;
total income and expenses;
the first 10 rows (…and N more), with 🆕 on tags that will be created and ~ on suggested tags;
what your message changed, and how tags were mapped;
warnings: rows that were skipped and why, rows in more than one file (imported once), rows that are already saved (same date, amount, and note), more than one currency (only USD keeps one), ambiguous dates, or a file where every amount is positive (flip the signs fixes it).

Rows that are already saved are skipped, so sending the same file twice doesn’t add anything twice. Nothing is saved until you press Confirm. Then all rows (and new tags) are saved together: either all of them or, if something fails, none. Transactions get the Web or Telegram source, like other assistant changes. Amounts are saved as they are, with no currency conversion.

CSV and Excel layout

Download a template: CSV, JSON, OFX, or QIF.

date,amount,note,tags
2026-10-01,-350.50,Silpo,Groceries
2026-10-02,-120,Coffee with Anna,Coffee
2026-10-05,25000,Salary,
2026-10-07,-899.99,Netflix and Spotify,Subscriptions|Entertainment

The same in JSON (any list of objects works; the key names follow the CSV rules below):

{
  "transactions": [
    { "date": "2026-10-01", "amount": -350.50, "note": "Silpo", "tags": ["Groceries"] },
    { "date": "2026-10-05", "amount": 25000, "note": "Salary" }
  ]
}

Required: a date column and an amount column. Expenses are negative and income is positive. Instead of one amount you can have two columns, such as Debit / Credit, or Expense / Income. A Type column with debit / credit or expense / income sets the sign too.
Optional: a note (description, merchant, purpose) and tags. Separate several tags with |, ,, or ;.
Headers can be in English, Ukrainian, or Russian, and the header row doesn’t have to be the first one. Bank exports such as Monobank or PrivatBank usually work as they are. If the columns can’t be recognised, the AI looks at the header and 3 sample rows to work them out.
Separators: ,, ;, tab, or |. Decimal comma (-350,50), spaces or apostrophes in thousands, and amounts in parentheses (350.00) all work. UTF-8 and Windows-1251 files work.
Dates: 2026-10-01, 01.10.2026, 01/10/2026, 20261001, 1 Oct 2026, or real Excel dates. When a file only has dates like 03/04/2026 that could be either way, the AI decides from the file if it’s reading it anyway; otherwise your date format in Settings is used and the card warns you. Write dates are month first (or day first) to set it. QIF files are read month first unless the dates show otherwise.
Up to 1000 rows per import, from all files together. A longer import gets a warning; split it into shorter periods.

AI assists

CSV, Excel, OFX, QIF, JSON, or an HTML table whose columns are recognised: free, no AI assist used.
A spreadsheet whose columns need the AI, any PDF or Word file, or an HTML page without a table: one AI assist per file.
Screenshots: one AI assist for all of them.
Your message with the files: free for common phrases or when the import already used an assist, otherwise one assist. Replies to the tags question work the same way.
An assist is counted only when the AI actually read something. When none are left, a PDF receipt asks for the amount instead (the same as a photo), and files with recognised columns still work.

Very long PDF statements may not fit in one AI read. If rows are missing from the preview, export the statement as CSV, Excel, or OFX instead.

Privacy

Files are stored privately like receipt photos and deleted once they’re read, or when the conversation expires. Only a PDF receipt sent on its own is kept, attached to its transaction. The text inside a file is treated as data: instructions written in a document are ignored. Only your own message can change the import.