Document parsing APIThe extraction step, without the extraction problem
Inteleto's document parsing turns PDFs, scans, images and office files into clean Markdown plus positioned blocks — each one carrying its page number and position on the page, so anything you build on top can cite back to the exact spot in the original. Pages that already have a text layer are read on CPU; only scanned and image pages go to the vision model, which is what keeps the cost of a mixed archive sane.
Built because we needed it ourselves and the paid alternatives were costing five figures a month. It runs in production on scanned lending files, contracts and technical manuals.
Private beta · no card · we reply by hand
Five stages, one contract
A file goes in and Markdown, positioned blocks and retrieval-ready chunks come out. Each stage is separately versioned, so a change to the OCR engine or the chunking rules re-runs only the work it invalidated instead of the whole corpus.
- 1Validate
What the file really is, how many pages it has, and which route it needs. A .pdf that is actually a TIFF is caught here rather than three stages later.
- 2Split
Documents are cut into fixed 64-page parts. A part is the unit of work, so a 900-page file cannot hold the queue while two-page uploads wait behind it.
- 3Extract
Pages with a usable text layer are read on CPU. Scanned and image pages are rasterized at 200 DPI and read by a vision OCR model. The decision is per page, not per document.
- 4Normalize
Tables become Markdown pipe tables with merged cells flattened into a grid, inline style noise is stripped, and blocks are stitched back into one document with their page numbers rebased.
- 5Chunk
Structure-aware parent/child chunks, token-capped, each prefixed with its heading path. Deterministic — no model calls in the ingest path.
What it measures at
One 12GB GPU, 44-page scanned underwriting manual.
Vision output against the same PDF's own text layer.
Measurably weaker than prose. Digits are where OCR fails.
These are our own measurements on our own hardware, published so you can argue with them rather than take them on faith. Accuracy is measured against the embedded text layer of digital PDFs, which is a proxy for ground truth and not the same thing as it. Ask us to run your documents and you get numbers for your documents.
Why the per-page routing matters
Roughly 60% of pages in a typical mixed archive are born-digital: they already contain their own text, perfectly, for free. Sending those through a vision model is the single largest avoidable cost in document parsing, and it is what most per-page pricing quietly charges you for.
Extraction is also content-addressed per page range and keyed to the engine version, so a range that has already been extracted at the current version never reaches a GPU again — no matter who uploads the file. Re-processing a corpus after a chunking change costs nothing in extraction.
Want the whole thing, not just the parser?
The parsing stage is the front of a platform that indexes, retrieves and answers with a citation on every claim. See what it is built into.
Frequently asked questions
- What do I get back from a parse?
- Markdown for the document, plus the blocks it was assembled from. Every block carries its type, its page number and its position on that page, so a downstream system can highlight the exact region a claim came from instead of citing a whole file. Tables come back as Markdown pipe tables, not as HTML with inline styling.
- How does it handle scanned pages versus digital ones?
- Per page. A page with a usable text layer is read directly on CPU, which is near-instant and costs nothing in GPU time. A scanned or image-only page is rasterized at a controlled 200 DPI and read by a vision OCR model. A document that mixes both — a signed contract with scanned exhibits, say — is handled page by page rather than being forced down one route.
- How fast is it?
- 0.31 seconds per page in steady state on a single 12GB GPU, measured on a 44-page scanned underwriting manual. Documents are cut into 64-page parts and worked as parts, so a 900-page file does not sit in front of a two-page one.
- How accurate is the OCR?
- On the internal benchmark that compares vision output against the same PDF's own embedded text layer, word-level F1 is 0.975 and numeric recall is 0.872. Numeric recall is the number to watch: digits are where OCR fails in ways that matter, and it is measurably weaker than prose.
- Can I use the parser without the rest of the platform?
- That is what this is for. The parsing layer is a separate stage with its own contract, and standalone access is in private beta — request access below and we will tell you what your documents cost to run before you commit to anything.