Document OCR

Turn a digitally rendered page into Markdown in one chat completion: headings, tables, equations, reading order.

Transcribe a page

Send one chat completion per page with the Quickstart client: image plus prompt, temperature=0.0, thinking off.

Python

OCR_PROMPT = (
    "Transcribe this document page to Markdown, preserving the reading order, "
    "headings, lists, and tables. Render tables as Markdown tables and equations "
    "as LaTeX. Return only the transcription, with no commentary and no surrounding "
    "code fence. If the page has no readable text, return an empty string."
)

def ocr_page(image_url: str) -> str:
    response = client.chat.completions.create(
        model="holo4-27b",
        messages=[{
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": image_url}},
                {"type": "text", "text": OCR_PROMPT},
            ],
        }],
        temperature=0.0,
        extra_body={"chat_template_kwargs": {"enable_thinking": False}},
    )
    return response.choices[0].message.content or ""

print(ocr_page("https://your-host/page.png"))  # or "data:image/png;base64,..."

Input	Quality
English, digitally generated: exported PDFs, slides, web pages, reports	Strongest
Scanned pages and photos	Best effort
Handwriting, non-Latin scripts	Not a good fit. Use a dedicated OCR system when stakes are high

Multi-page PDFs

Holo reads images, not PDFs, so rasterize each page to an image and transcribe them one per request with ocr_page, then stitch the results. One page per request keeps each image at full resolution and is the most reliable pattern.

Python

import base64
import pymupdf  # pip install pymupdf

pages = []
with pymupdf.open("document.pdf") as doc:
    for page in doc:
        png = page.get_pixmap(dpi=200).tobytes("png")
        pages.append(ocr_page("data:image/png;base64," + base64.b64encode(png).decode()))

markdown = "\n\n".join(pages)
print(markdown)

Rasterize at 200 to 300 DPI (scale: 2 to 3). Lower resolution loses small text. Above that the server downsizes the image, so extra DPI only costs upload bandwidth. Run pages concurrently to speed up long documents, within your rate limit.

Output is capped at the model’s max_output_length in GET /v1/models, and a dense page (large tables, small print) can exceed that. Check finish_reason and treat length as a truncated transcription. Split dense pages into two images.