Thai PDF to Markdown

Thai PDF to Markdown with the PDF's exact text. Thai PDFs whose text is secretly broken are caught and re-read. Ready for RAG and AI agents.

แปลง PDF ภาษาไทยเป็น Markdown ที่สะอาด โดยใช้ข้อความจริงจาก PDF พร้อม OCR ภาษาไทย สำหรับงาน RAG และ AI

• The PDF's own text where it exists. Many Thai PDFs already contain their text. It is copied from the PDF, with no AI rewriting, in about 2 seconds per page. Headings, tables and reading order are rebuilt from the page layout.
• Broken text is caught. Many Thai PDFs look right on screen but hold the wrong text: legacy Thai fonts, wrong encodings, and the TH SarabunPSK bug where "การ" comes out as "กำร". A page that shows these faults is read by OCR instead. Not every broken page is caught: see Good to know.
• Thai OCR only where it's needed. Scanned pages are read with a Thai OCR model fine-tuned on Thai government documents. See Accuracy.
• Private. Documents are processed in memory, never stored or viewed by us, and never used for training.
• Built for pipelines. PDF links in, Markdown out: API, Python, n8n, Make, Zapier or AI agents (Apify MCP server).

See it on real pages: four Thai government PDF pages, before and after, each shown as printed, as stored inside the PDF, and as this Actor returns it.

Quick start

Paste PDF links into PDF URLs, click Start, then open the Output tab. The form starts at 5 pages per PDF so a first test costs almost nothing; clear Max pages per PDF to convert whole files.

Field	Type	Default	Notes
urls	array of strings	—	HTTPS links to PDF files. Up to 100 links per run; the same file linked twice is converted once.
maxPages	integer	1000	Optional. Each PDF is converted whole, up to 1,000 pages; set a lower number to convert only the first N pages.

One request, for a few pages (a run longer than 300 seconds returns HTTP 408):

curl -X POST "https://api.apify.com/v2/acts/vetsuyan-labs~thai-pdf-to-markdown/run-sync-get-dataset-items" \
  -H "Authorization: Bearer $APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://example.com/report.pdf"], "maxPages": 5}'

Python, for whole documents, with apify-client 3 or newer:

from datetime import timedelta

from apify_client import ApifyClient

client = ApifyClient("<APIFY_TOKEN>")
run = client.actor("vetsuyan-labs/thai-pdf-to-markdown").call(
    run_input={"urls": ["https://example.com/report.pdf"]},
    run_timeout=timedelta(hours=1),  # 200 scanned pages take about 17 minutes
)

# One item per page
for page in client.dataset(run.default_dataset_id).iterate_items():
    print(page["page"], page["source"], page["error"] or page["markdown"][:80])

# Each document as one Markdown file
store = client.key_value_store(run.default_key_value_store_id)
for doc in store.get_record("OUTPUT")["value"]:
    if doc["markdownKey"]:
        markdown = store.get_record(doc["markdownKey"])["value"]
    if doc["error"] or doc["pagesSkipped"] or doc["pagesFailed"]:
        print(doc["sourceUrl"], "is incomplete:", doc["error"] or "pages are missing")

Output

One dataset item per page:

{
  "sourceUrl": "https://example.com/report.pdf",
  "page": 3,
  "source": "text_layer",
  "markdown": "## ตารางที่ 1 ...",
  "avgConfidence": null,
  "processingMs": 1650,
  "error": null
}

• source is text_layer (copied from the PDF) or ocr. avgConfidence is set for OCR pages only.
• A blank page has empty markdown. A page that cannot be converted has an error, and the rest of the document continues. A file that cannot be processed (not a PDF, too large, password-protected, download blocked) is one item with page: null and an error.

In the key-value store:

• document-1.md, document-2.md, … hold each file's full Markdown, in the order of your links (markdownKey in OUTPUT names each). Pages are separated by <!-- page N --> comments.
• OUTPUT lists, for each link the run reached: sourceUrl, markdownKey, pagesConverted, pagesTotal, pagesSkipped, pagesFailed, error and stoppedByLimit.
• Check pagesSkipped and pagesFailed. If either is above 0, the Markdown is not the whole file. Skipped pages were beyond 1,000 pages, your maxPages or your spending limit (stoppedByLimit). A failed page leaves a <!-- page N: could not be converted --> comment in its place.

Accuracy

Digital PDFs. On pages reported as text_layer, the text is the PDF's own. No AI model reads or rewrites it, so no word is invented. It is only tidied: spacing, line wrapping, the order and encoding of Thai vowel and tone marks, invisible characters, and letters a PDF draws twice. What can still go wrong is the structure: a heading that is not marked as one, a table that comes out as plain text, or a line taken for a page header or footer and left out.

Scanned pages are read by OCR, and OCR makes mistakes. We ran this Actor on ThaiOCRBench, a public benchmark of real Thai images. Scores run from 0 to 1; higher is better.

What was tested	This Actor	Gemini 2.5 Pro	Claude Sonnet 4	GPT-4o	Qwen2.5-VL 72B
Reading a full printed page (197 images)	0.786	0.897	0.661	0.610	0.720
Reading a table (193 images)	0.444	0.686	0.650	0.571	0.549
Reading handwriting (209 images)	0.156	0.714	0.301	0.489	0.393

• Printed pages: a good fit. Expect some wrong characters on every scanned page, more on old or faint ones.
• Tables in scans: simple tables work, but the large models do better. Check any figures you rely on.
• Handwriting: not what this Actor is for.

The other models' scores were published by the benchmark's authors in 2025 (paper); we did not test those models, and their newer versions may do better. Ours were measured in October 2026. Some table images were used while we developed this Actor; on the 129 that were not, it scores 0.414.

Pricing

Pay per converted page. No subscription.

Event	Price	Per 1,000 pages
Page from the PDF's text layer	$0.003	$3
Page that needed Thai OCR	$0.010	$10

• A page counts as OCR when it was read whole by OCR (a scanned page, a page whose own text is broken, or a page with very little text beside a picture, such as many cover pages), or when OCR read 20 or more letters and digits in the pictures or charts of a digital page. Many digital documents contain a few OCR pages.
• Example: a 100-page report with 70 digital and 30 scanned pages costs $0.51.
• Blank pages, failed pages and failed documents are free. Apify's start fee for a run (about $0.0002) may also apply.
• Apify's free plan ($5 a month) covers about 1,600 text pages or 500 OCR pages.
• Set a maximum cost per run in Apify and the Actor stops cleanly when it is reached. A new run starts every PDF from page 1 and charges those pages again, so set the limit to cover the whole job.

Privacy and security

• We don't keep or look at your documents. They are processed in memory and never written to disk. Results (your links, the converted text and a progress record) go only to your own Apify storage and are deleted according to your plan's data retention, or at once when you delete the run.
• Never used for training. Your documents and results do not train or improve any model.
• A run you share with us (Apify's "Share run data with developers" setting, or a run attached to an issue report) is used only to answer your request. We keep no copies.
• Limited permissions, private logs. The Actor uses only this run's storage. Logs show each link's host name, page counts and error types: never the path, the query string or any document text.
• Safe downloads. HTTPS only, with certificates verified. Private and internal network addresses are refused.
• For sensitive documents, pass short-lived signed links that outlive the run timeout you set.
• Treat the output as untrusted content. The Markdown is the document's own text. It can contain text written to manipulate AI models (prompt injection), HTML or links: apply your usual safeguards before passing it to an AI agent or showing it as a web page.

Personal data (Thailand PDPA). If your documents contain personal data, you are the data controller and we are your data processor: we process documents only to convert them as you request, keep them only in memory while the run lasts, and never share them. Processing runs on Apify's platform, hosted outside Thailand, so it involves a transfer abroad: you are responsible for having a basis for it (PDPA sections 28–29), and Apify's Data Processing Addendum covers Apify's part as our sub-processor. If we become aware of a security incident affecting this Actor we will publish a notice on this page and notify Apify without undue delay. For data requests or privacy questions, email the contact address on our Apify profile.

Good to know

• Headers, footers and seals are removed: page numbers, running headers and official stamps. Now and then the first or last line of a page is taken for a header or footer and left out, and on a scan, text touching a stamp can go with it.
• Scans: old or low-quality scans contain more wrong characters. A line that cannot be read reliably is left out rather than guessed.
• Scanned PDFs that are already searchable: if you can select the text in a scanned PDF, another program has already added its own reading of each page. This Actor normally uses that text (source is then text_layer, at the lower price), so any mistakes in it come through. Check a few pages first.
• The Thai font bug, the other way round: a PDF where "สำ" is exported as "สา" ("สาหรับ", "อาเภอ") usually passes through as written.
• Tables: tables with ruling lines become Markdown tables. On digital pages, a layout that cannot be read reliably as a table is output as plain text in reading order, not as a wrong table.
• Pictures, charts and forms: text inside pictures and charts is read by OCR and may come out fragmented; a label shorter than 20 letters and digits is left out. Values typed into PDF form fields are included.
• Files the Actor cannot fetch: some sites block automated downloads. For those and for local files, pass a link you control that works without logging in, such as a pre-signed S3 or Google Cloud Storage URL.
• Limits: 100 MB per file, 5 minutes per download, at most 3 redirects, and standard HTTPS links only. Of a file longer than 1,000 pages the first 1,000 are converted; if you need more, tell us in the Issues tab.
• Large jobs: a page takes about 2 seconds from the PDF's own text and about 4–6 seconds with OCR, so 1,000 scanned pages cost $10 and can take an hour and a half. Set the run timeout to match. If a run times out, resurrect it with a longer timeout: it continues where it stopped and never charges a page twice.

ภาษาไทย

แปลงไฟล์ PDF ภาษาไทย (รายงาน ระเบียบ ราชกิจจานุเบกษา คู่มือ เอกสารสแกน) เป็น Markdown ที่สะอาด พร้อมใช้กับ RAG และ AI

• แม่นยำเมื่อ PDF มีข้อความอยู่แล้ว: คัดลอกข้อความจริงจาก PDF โดยไม่ใช้ AI เขียนใหม่ พร้อมหัวข้อ ตาราง และลำดับการอ่าน
• เข้าใจปัญหา PDF ภาษาไทย: ตรวจจับฟอนต์ไทยที่เสียโดยอัตโนมัติ เช่น "การ" ที่กลายเป็น "กำร" แล้วอ่านหน้านั้นด้วย OCR แทน
• OCR ภาษาไทยเฉพาะหน้าที่จำเป็น: หน้าที่เป็นภาพสแกนจะอ่านด้วยโมเดล OCR ภาษาไทยที่ปรับแต่งจากเอกสารราชการจริง
• เป็นส่วนตัว: ประมวลผลในหน่วยความจำ เราไม่เก็บหรือเปิดดูเอกสารของคุณ ไม่นำไปฝึกโมเดล และผลลัพธ์อยู่ในพื้นที่จัดเก็บ Apify ของคุณเองเท่านั้น
• ราคา: จ่ายตามจำนวนหน้า หน้าว่างและไฟล์ที่ผิดพลาดไม่คิดเงิน
• ดูตัวอย่างจริง: เอกสารราชการ 4 หน้า ก่อนและหลังแปลง