Ukrainian scanned PDFs to Markdown for RAG and LLMs

Andriy Kuzmenko — Sat, 07 Nov 2026 00:31:17 +0000

Most document-to-text tools are tuned for English. A scanned Ukrainian invoice or court decision often comes back
as Latin look-alike gibberish, or with its tables lost. This guide shows a pipeline that keeps Cyrillic text,
headings and tables, and shows exactly what it does and does not do - every output below is from a real run on
6 November 2026.

Update, 7 November 2026. The first run showed two problems: a table in a scan came back as one flat row, and
["ukr", "eng"] turned «Арматура» into Apmatypa. Both are fixed now; the outputs below are from the fixed
version (build 0.1.14), and the old ones are kept in the text so you can see the difference.

Disclosure: I built the tool used here, Document to Markdown for AI
on Apify. Its source code is open: AndriyVK/-apify-document-tools.

What you get

One record per document: the Markdown, every table as rows of cells, optional chunks with their nearest heading
(ready for a vector database), and page counts. PDF pages with a text layer are read directly; scanned pages and
images go through Tesseract OCR, which has a proper Ukrainian model.

Formats: PDF, scanned PDF, PNG/JPG/TIFF, DOCX/DOC, XLSX/XLS, PPTX/PPT, ODT/ODS/ODP, RTF, HTML.

Run it from Python

pip install apify-client

from apify_client import ApifyClient

client = ApifyClient("<YOUR_APIFY_TOKEN>")
run = client.actor("ai_admin/document-to-markdown").call(run_input={
    "sources": [
        {"url": "https://api.apify.com/v2/key-value-stores/FloomuI2DrQCJazY8/records/scan-uk.pdf"},
        {"url": "https://api.apify.com/v2/key-value-stores/FloomuI2DrQCJazY8/records/sheet-uk.xlsx"},
    ],
    "ocrMode": "auto",          # OCR only pages without a text layer
    "ocrLanguages": ["ukr"],    # list only the languages that are really in the scan
    "includeTables": True,
    "chunkSize": 800,           # characters per chunk, 0 = no chunks
})
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item["fileName"], item["ocrPages"], "OCR pages")
    print(item["markdown"])

The two sample files are public, so you can run this as is. Your own files work by URL, or from your Apify
key-value store.

The real output

A one-page scan (scan-uk.pdf, no text layer - 1 OCR page):

Звіт про постачання

Склад у Львові отримав 1240 палет будівельних матеріалів протягом

третього кварталу. Терміни доставки скоротилися на дванадцять відсотків після запровадження

нового графіка.

| Товар | Кількість | Ціна |
| --- | --- | --- |
| Цемент | 420 | 96,50 |
| Арматура | 180 | 412,00 |

The same content as an Excel file (sheet-uk.xlsx, 0.2 s) - here the structure is exact:

## Data

| Товар | Кількість | Ціна |
| --- | --- | --- |
| Цемент | 420 | 96,50 |
| Арматура | 180 | 412,00 |
| Total | 600 | 1,234.50 |

Cell values are kept as written: 96,50 stays 96,50, it is not turned into a float.

One setting that matters: OCR languages

On 6 November, with "ocrLanguages": ["ukr", "eng"], the word «Арматура» came back as Apmatypa - the OCR picked
Latin letters that look the same. Now such a word is put back in Cyrillic when it is in Tesseract's Ukrainian
dictionary and the Latin spelling is not an English word, so Excel, API or copy stay as they are. The run on
7 November with ["ukr", "eng"] gave «Арматура» correctly. Still, list only the languages that are really on the
page: every extra language is one more way for the OCR to guess wrong.

Price

You pay per page: $1 per 1,000 text pages and $4 per 1,000 OCR pages (Apify pay-per-event pricing; OCR is
charged only for pages that needed it). A document that cannot be downloaded or read is not charged. The two files
above cost a fraction of a cent.

Limits - honestly

Tables in scans come back as tables when they have ruled grid lines and the scan is clean. A borderless table, or a grid lost in a noisy 200 dpi scan, still comes out as text. If you have the original Word or Excel file, convert that instead - its tables are exact.
OCR quality follows the scan. Clean 300 dpi scans read well; phone photos at an angle are worse. For photos, the Image to Text OCR Actor auto-rotates pages and returns a confidence for every word.
Handwriting is not supported.

Use it from an AI agent

The same tool is available through the Apify MCP server, so Claude, Cursor or any MCP client can call it:
https://mcp.apify.com?tools=ai_admin/document-to-markdown. A local MCP server with all our tools is in
the GitHub repository.

Originally published at andriyvk.github.io.

DEV Community: Andriy Kuzmenko