Document
extraction
Methods for extracting text, tables and data from PDFs and scanned documents, and how to choose between them.
Press → to start · O for overview · N for notes
How PDF libraries reconstruct structure
- The characters stored in the file, with exact positions, and no new recognition step
- Milliseconds per page on a laptop, so millions of pages cost almost nothing
- Completely repeatable and easy to audit
- Scans with no text layer, or a poor hidden OCR layer
- Text layers that are garbled, incomplete or wrong
- Columns, headers and footnotes in the wrong order
- Tables without ruling lines, and tables that span pages
- Maths, and files with broken font maps
olmOCR-Bench figures from the Marker README (Datalab, vendor-run)
Cloud document services and pricing
Azure, AWS and Google offer similar services. Each layer is a separately trained model, and each costs more per page.
List prices per 1,000 pages, first volume tier, checked 22 Oct 2026. Azure Retail Prices API, AWS Price List, Google Document AI pricing page.
- Boxes and confidence on everything
- Runs inside the client's own cloud
- Mature, with service agreements
- Complex financial tables, charts, maths
- Output is a large JSON graph to walk
- Costs add up as you add layers
Two influential research results
Using word positions as model input
A number to the right of "Total" is probably the total. Feeding positions into a language model made form and invoice extraction much better.
Generating output directly from the image
This switched the task from recognising characters to generating text. It handles difficult pages better, but it can also produce text that is not on the page.
Three designs of specialist model
A pipeline of small models
Separate models for layout, tables, formulas and OCR, joined by code. Uses the PDF's own text where it can.
Layout first, then read each block
Find the blocks on a small, low-resolution copy of the page. Then read each block at full resolution.
A single model converts the whole page
A small vision language model (1–9 billion parameters) takes the page image and outputs the full page as markdown or HTML, often with block positions.
Images: IBM Docling (MIT), OpenDataLab MinerU2.5 (arXiv 2509.22186), Datalab Chandra 2 blog. Click to zoom.
The stages of a document pipeline
Images: Datalab Surya README (GitHub). Surya 2 (May 2026) now does layout, OCR and tables in one 650M-parameter model, and gives each block a confidence score.
Reducing the number of tokens per page
DeepSeek-OCR (Oct 2025) tested how few tokens are needed to represent a page.
Compare a general model, which uses 560 to 2,700 image tokens per page. Fewer tokens reduce cost and processing time. Together with smaller models, this explains why self-hosted costs are now $0.10–0.70 per 1,000 pages in GPU time.
Strengths of generative models
"axes": "x: year; y: salary ($)", "data": "<table>… 2010 | ~66,000 …", "description": "Average salary fell, then recovered…"
Chart data extraction
Values are read off the axes, marked "~" when estimated. OCR can only return the axis labels. Example from Chandra 2.5, a specialist model; general models such as Gemini and Claude handle the same cases. Image: Datalab blog.
Handwriting and forms
Handwriting, ticked boxes, stamps and photographs of documents. Training on large and varied data allows these models to read partly legible words. Example from Chandra, a specialist model. Image: Datalab README.
{
"party_a": "Acme Ltd",
"party_b": "Example Services LLC",
"notice_period_days": 90,
"payment_terms_days": 30,
"governing_law": null,
"auto_renews": null
}
// null = not stated on the page.
// Never default a missing value to false.
Direct extraction of specified fields
Fields from the slide 7 contract. Use the provider's structured-output mode with a schema: the output then matches the schema's format, but values can still be wrong and responses can be refused or cut off. Validate before use.
Weaknesses of general models
Datalab anti-hallucination blog, vendor-run
| What happens | Why |
|---|---|
| It corrects the text | The model favours likely words, so it corrects typos, normalises dates and fills gaps. |
| It leaves things out | It hits its output limit, skips repeated rows or summarises. The request still succeeds: check the finish reason and compare row counts. |
| It repeats itself | It can loop on dense or unusual pages. |
| It misreads small print | The page image is reduced to fit the model's size limit, which can make small text illegible. |
| Errors on long documents | When many pages are sent in one request, content from different pages can be mixed up. Parse page by page, then combine the results and reconcile values that span pages. |
| No confidence scores | The output does not show which parts are uncertain, so reviewers cannot tell where to check. If asked for word or block locations, the model generates approximate coordinates rather than measuring them from the page. |
| Results change | Providers update models under the same name, and outputs can differ between runs even at temperature 0. |
Verification features in recent tools
- Use the PDF's own text when it is good. Only read pixels where it is missing or broken. (Marker, Docling, Gemini and Claude with PDF input)
- Detect layout at low resolution and read text at full resolution. (MinerU, PaddleOCR-VL)
- Small model to transcribe, large model only for meaning. (Marker with an LLM, Google Layout Parser)
- Check, retry and flag uncertainty. Re-read doubtful blocks and mark text "[illegible]" instead of guessing. (Datalab)
- Return source positions. Mistral OCR 4, Chandra and Docling all return boxes, so a reviewer can see where each value came from.