pagemark

pagemark converts PDFs to Markdown with vision-language models.

Use it with any OpenAI-compatible endpoint. You provide the base_url, model, and api_key. pagemark renders each page, grounds the model with the PDF text layer, and checks the result for common failures: repetition loops, truncation, dropped columns, and hallucinated tables. If a check fails, it retries with a stricter strategy and falls back to the text layer if the model still cannot recover.

If one page fails, the document still finishes. pagemark records warnings for that page instead of aborting the whole conversion.

Prerequisites

• Python 3.10+
• An OpenAI-compatible endpoint serving a vision-language model (Ollama, vLLM, LM Studio, any hosted provider, etc.)

Install

pip install pagemark            # library only
pip install "pagemark[cli]"     # with CLI

Quick start

from pagemark import Pagemark

client = Pagemark(
    base_url="http://localhost:11434/v1", # OpenAI-compatible API
    api_key="your-key", # API key
)

doc = client.convert("paper.pdf", model="qwen3-vl:8b")

print(doc.markdown)
doc.save("out.md", assets_dir="out_assets/")

api_key is optional. pagemark resolves it from the explicit argument first, then from OPENAI_API_KEY. If neither exists, it raises an error that names the base_url.

Async

aconvert() is the implementation behind convert(). convert() wraps it with asyncio.run(). Pages run concurrently, bounded by a semaphore. The default limit is 4.

from pagemark import Pagemark

client = Pagemark(
    base_url="http://localhost:11434/v1", # OpenAI-compatible API
    api_key="your-key", # API key
)

doc = await client.aconvert(
    "paper.pdf",
    model="qwen3-vl:8b",
    concurrency=8,
)

Page selection

Convert selected pages instead of the whole document:

doc = client.convert("paper.pdf", model="qwen3-vl:8b", pages="1-5,9,12-")

pages accepts "3", "1-5", "1-5,9,12-", or a list[int]. Out-of-range indices raise an error instead of being silently clipped.

Bring your own model

If you need settings the constructor does not expose, pass a pre-configured BaseChatModel. That covers custom timeouts, proxies, non-OpenAI LangChain backends like ChatAnthropic or ChatOllama, callbacks, and similar options.

from langchain_openai import ChatOpenAI
from pagemark import Pagemark

my_model = ChatOpenAI(
    model="...",
    base_url="...",
    api_key="...",
    timeout=120,
    max_retries=5,
)

client = Pagemark(chat_model=my_model)
doc = client.convert("paper.pdf")

CLI

pagemark convert paper.pdf -o paper.md \
    --model qwen3-vl:8b \
    --base-url http://localhost:11434/v1

pagemark convert paper.pdf -o paper.md --json paper.json \
    --model qwen3-vl:8b \
    --base-url http://localhost:11434/v1

pagemark doctor --base-url http://localhost:11434/v1 --model qwen3-vl:8b

doctor checks whether the endpoint is reachable, whether the model is present, and where the API key came from.

How it works

pdf -> render + anchor -> prompt (profile) -> ChatOpenAI -> guardrails -> parse -> assemble -> Document
       pypdfium2          per-model           langchain      retry ladder   md-it    stitching   md + json

1. Render each page to a PNG with pypdfium2, downscaled to the profile's pixel budget.
2. Extract the PDF text layer as an anchor for the model. Long pages are middle-elided to about 6k characters.
3. Send the image and anchor text to the VLM.
4. Check every response with five guardrails.
5. Retry with stricter strategies when a guardrail fails.
6. Assemble the pages into a Document with typed blocks, then stitch across pages by stripping running headers and footers, joining split paragraphs, and dehyphenating text.

Guardrails

Escalation ladder

When any guardrail fails, pagemark retries with more aggressive strategies:

Each page includes source, confidence, and warnings, so downstream consumers can filter by quality.

Output

convert() returns a Document with:

• markdown: the full stitched Markdown output
• pages: list of Page objects, each with:
  • markdown: per-page Markdown
  • blocks: typed blocks (heading, paragraph, table, figure, code, list, formula, caption, footnote)
  • source: "vlm" or "text_layer" (if the model couldn't handle the page)
  • confidence: 0-1 score
  • warnings: what went wrong, if anything
• assets: embedded images extracted from the PDF
• model_dump_json(): full structured output as JSON
• save(path, assets_dir=...): writes the Markdown file and extracted images

License

MIT. Runtime dependencies are permissively licensed - mostly MIT/Apache-2.0/BSD, with a few under MPL-2.0 (certifi, orjson) and PSF-2.0 (typing-extensions). No GPL/LGPL/AGPL dependencies in the tree.