IntraParse
The hosted way to use Meridian. Drop in a PDF or call the API and you get back tidy, structured data. No GPU, no infrastructure, no queue. Built for developers who want the parsing without running the pipeline. On a 69-page scanned NASA report it finished in 59 seconds against LlamaParse’s 15+ minutes, and returned all 69 pages where LlamaParse managed 8.
- 59 sec
- vs 15+ minutes
- 69 of 69
- pages, vs 8 of 69
- ~1/5th
- the cost
What the problem was
Meridian will parse anything you throw at it, but only if you have a GPU and the patience to stand up Docling, vLLM, Ollama, Redis, Celery and Qdrant and keep them talking to each other.
Most developers who need good parsing don’t want any of that. They want an endpoint, a key, and a shape of JSON they can rely on.
The alternative they usually reach for is LlamaParse, which bills per page to call someone else’s API. On a 69-page scanned NASA report it took over fifteen minutes and returned eight pages. IntraParse took 59 seconds and returned all sixty-nine.
How I built it
One endpoint, structured data back
Drag a PDF into the browser or send it to the API with a key, and it is the same pipeline either way. What comes back is one of four things per chunk: text, a table, a figure or a formula. That’s the entire contract.
There’s an async endpoint for long documents, full history so you can re-download anything you’ve parsed, and a side-by-side view in the browser to check the output against the original page.
Open models are why it costs a fifth as much
The commercial parsers bill you per page to call someone else’s API, and that price never comes down no matter how much you send them. Meridian runs on open models on hardware I control, so the cost is compute rather than a licence: roughly a fifth of what the hosted tools charge for the same document, and faster besides.
Cheap enough to put a whole archive through, rather than only the pages you can justify.
How it turned out
- 59 seconds against LlamaParse’s 15+ minutes on the same scanned NASA report, and all 69 pages instead of 8.
- Free while in beta, with an API, a dashboard and full document history.
- If you’d rather run it yourself, the engine underneath is open source.