Extract Text from PDF Documents

Vision models on the IONOS CLOUD AI Model Hub accept image input only. To process a PDF document, each page must be rendered as an image before it is sent to the model. This guide shows how to do that using pypdfium2, a lightweight PDF rendering library, together with the OpenAI-compatible API.

About this guide

In this guide, you will learn how to:

• Render PDF pages as images using pypdfium2
• Extract text from a PDF using LightOnOCR-2-1B, a model optimized for document OCR
• Extract and understand content from a PDF using Mistral Small 24B, a multimodal language model

Prerequisites

Before you start, make sure you have:

• Python 3.11 or higher installed on your machine
• The IONOS_API_TOKEN environment variable set with your authentication token
• A PDF file to process

Download the code files to follow with ready-to-use examples:

Python Notebook

Download the Python Notebook to explore PDF text extraction with ready-to-use examples.

ai-model-hub-pdf-to-text.ipynb

6KB

Open

Python Code

Download the standalone Python script for a quick implementation.

ai-model-hub-pdf-to-text.py

4KB

Open

Step 1: Install dependencies

pip

pip install pypdfium2 pillow openai

uv

uv add pypdfium2 pillow openai

Step 2: Choose a model

The IONOS CLOUD AI Model Hub offers two model types suitable for PDF text extraction:

Model	Identifier	Best for
LightOnOCR-2-1B	lightonai/LightOnOCR-2-1B	Pure text extraction from scanned documents and complex layouts
Mistral Small 24B	mistralai/Mistral-Small-24B-Instruct	Text extraction combined with document understanding and Q&A

Step 3: Render a PDF page as an image

pypdfium2 renders each PDF page as a PNG image in memory. The scale=2.0 parameter renders at 200 DPI, which balances text sharpness with request payload size.

import base64
import io

import pypdfium2 as pdfium


def pdf_page_to_base64(pdf_path: str, page_index: int, scale: float = 2.0) -> str:
    doc = pdfium.PdfDocument(pdf_path)
    bitmap = doc[page_index].render(scale=scale)
    buf = io.BytesIO()
    bitmap.to_pil().save(buf, format="PNG")
    return base64.b64encode(buf.getvalue()).decode()

The function returns a base64-encoded PNG string ready to send directly to the API.

Request size limits: The total request body may not exceed 16 MiB, and each inline base64 data URI may not exceed 12 MiB after decoding. Both are refused with a 413. For high-resolution or dense pages, lower the scale value or split large documents into individual pages before sending.

Note: Concurrency limit on large requests applies. Requests larger than 4 MiB and requests sent with chunked transfer encoding are accepted two at a time. A further request is refused with a 503 and a Retry-After: 1 header. For more information, see Error Codes.

Step 4: Send a page to the API

LightOnOCR-2-1B

LightOnOCR-2-1B does not require a prompt — its output behaviour is embedded in the model weights.

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["IONOS_API_TOKEN"],
    base_url="https://openai.inference.de-txl.ionos.com/v1",
)

image_b64 = pdf_page_to_base64("document.pdf", page_index=0)

response = client.chat.completions.create(
    model="lightonai/LightOnOCR-2-1B",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/png;base64,{image_b64}"},
                }
            ],
        }
    ],
    max_tokens=4096,
    temperature=0.0,
)

print(response.choices[0].message.content)

Mistral Small 24B

Mistral Small 24B accepts a text prompt alongside the image, which lets you control the extraction or ask questions about the page content.

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["IONOS_API_TOKEN"],
    base_url="https://openai.inference.de-txl.ionos.com/v1",
)

image_b64 = pdf_page_to_base64("document.pdf", page_index=0)

response = client.chat.completions.create(
    model="mistralai/Mistral-Small-24B-Instruct",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "Extract all text from this document page and preserve the structure.",
                },
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/png;base64,{image_b64}"},
                },
            ],
        }
    ],
    max_tokens=4096,
    temperature=0.0,
)

print(response.choices[0].message.content)

Step 5: Process all pages in a PDF

To extract text from every page, iterate over the page count and collect the results:

import base64
import io
import os

import pypdfium2 as pdfium
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["IONOS_API_TOKEN"],
    base_url="https://openai.inference.de-txl.ionos.com/v1",
)


def pdf_page_to_base64(pdf_path: str, page_index: int, scale: float = 2.0) -> str:
    doc = pdfium.PdfDocument(pdf_path)
    bitmap = doc[page_index].render(scale=scale)
    buf = io.BytesIO()
    bitmap.to_pil().save(buf, format="PNG")
    return base64.b64encode(buf.getvalue()).decode()


def extract_pdf(pdf_path: str, model: str) -> str:
    doc = pdfium.PdfDocument(pdf_path)
    parts = []
    for i in range(len(doc)):
        image_b64 = pdf_page_to_base64(pdf_path, i)
        response = client.chat.completions.create(
            model=model,
            messages=[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "image_url",
                            "image_url": {"url": f"data:image/png;base64,{image_b64}"},
                        }
                    ],
                }
            ],
            max_tokens=4096,
            temperature=0.0,
        )
        parts.append(f"--- Page {i + 1} ---\n\n{response.choices[0].message.content}")
    return "\n\n".join(parts)


print(extract_pdf("document.pdf", model="lightonai/LightOnOCR-2-1B"))

What you learned

In this guide, you learned how to:

1. Render PDF pages as base64-encoded PNG images using pypdfium2
2. Send page images to LightOnOCR-2-1B or Mistral Small 24B through the OpenAI-compatible API
3. Process all pages in a document and collect the results

For pure OCR tasks, LightOnOCR-2-1B delivers fast and structured Markdown output. For tasks that combine extraction with document understanding, Mistral Small 24B offers more flexibility through its text prompt.

For more information about OCR capabilities, see the Optical Character Recognition (OCR) guide.