# Datalab

> Datalab turns any document into accurate, structured data. PDFs, scans,
> spreadsheets, and slides become markdown, HTML, JSON, or the exact fields
> you define. Datalab trains its own models (Chandra 2.6 runs the API),
> publishes open benchmarks, and runs in its cloud, your VPC, or air-gapped.

- Website: https://www.datalab.to
- Docs: https://documentation.datalab.to (index for agents: https://documentation.datalab.to/llms.txt)
- Quickstart: https://documentation.datalab.to/docs/welcome/quickstart.md
- OpenAPI: https://www.datalab.to/openapi.json
- Pricing (markdown): https://www.datalab.to/pricing.md
- Benchmarks (markdown): https://www.datalab.to/benchmarks.md
- Sign up: https://www.datalab.to/auth/sign_up (up to $20 of free usage each month)
- Agent setup: https://www.datalab.to/agent-setup.md (a coding agent can sign the person up with `datalab login` and set up the SDK)

## Quickstart

```python
# pip install datalab-python-sdk
from datalab_sdk import DatalabClient, ConvertOptions

client = DatalabClient()  # reads DATALAB_API_KEY

result = client.convert(
    "contract.pdf",
    options=ConvertOptions(output_format="markdown", mode="accurate"),
)
print(result.markdown)
```

## Processors and modes

- Convert: documents to markdown, HTML, JSON, or chunks. Modes: lite, fast, balanced, accurate. Lite returns text, tables and lists without images, add-ons or reusable checkpoints. The REST default is balanced; SDK and CLI default to fast.
- Extract: fields you define, as JSON. Modes: turbo, fast, balanced, accurate. Turbo has no citations or numeric confidence scores; fast supports asynchronous confidence scoring. Balanced and accurate add verification metadata. Block citations are the API default; phrase citations are open beta (SDK 0.9.2+) and cost nothing extra, with optional boxes when the value can be located. See https://documentation.datalab.to/docs/recipes/structured-extraction/api-overview.md.
- Segment: split a combined file into the separate documents it contains.
- Document agent (open beta): run an existing published processor with /api/v1/agent. Built-in processors include jats (JATS XML) and wcag (Accessible PDF). Check finished, verdict and verifier_metrics before using an artifact. See https://documentation.datalab.to/docs/recipes/pipelines/custom-processors.md.
- Form fill and form fields: fill PDF and image forms, or find a form's fillable fields and labels. Inspect detection in the signed-in playground: https://www.datalab.to/app/playground/form-fields.
- Redact (open beta): remove selected information from PDFs and images. Available PDFs can fail verification; inspect verdict and failing_checks before use. SDK helpers require 0.9.0+. See https://documentation.datalab.to/docs/recipes/redaction/redaction-api-overview.md.
- Track changes: tracked changes and margin comments from Word files, PDFs, and scans.
- Also: spreadsheet tables, create document, and add-ons for chart understanding, infographics, word bounding boxes, and table cell boxes.

## Prices (US dollars per 1,000 pages)

| Processor | What it does | Rate |
| --- | --- | --- |
| Convert — lite | Fast CPU OCR to markdown or HTML: text, tables and lists; no images or LLM passes | $0.50 |
| Convert — fast | Document to markdown or HTML, one model pass | $1.33 |
| Convert — balanced | Document to markdown or HTML, checked against the PDF text layer | $4 |
| Convert — accurate | Highest-fidelity conversion | $10 |
| Extraction — turbo / fast | Structured fields, single pass. Turbo returns JSON only; fast adds per-field citations and runs on Convert | $6 |
| Extraction — balanced | Structured fields, multi-pass with per-field verification | $15 |
| Extraction — accurate | Structured fields, multi-pass with per-field verification, strongest model | $20 |
| Segment | Detect boundaries between combined documents. Page-level split returns just the page ranges; block-level returns parsed markdown | $0.50 |
| Track changes | Diff revisions in a Word document | +$6 |
| Form fill | Populate a form's fields | $10 |
| Redaction | Remove information from a document by a redaction spec | $10 |
| Agent processor | Document agent workflows — the successor to custom processors, for higher-accuracy and more involved document work | $10 |
| Spreadsheet tables | Table extraction (500 cells count as one page) | $6 |
| Form fields | Find a form's fillable fields and their labels | $4 |
| OCR (legacy Surya) | Text-only recognition | $2 |
| Custom processor | Your own pipeline | $20 |
| Eval | Compare 2 to 4 versions of a pipeline on your documents, page by page or field by field | $20 |
| Chart understanding (add-on) | Adds chart parsing to a run | +$3 |
| Infographic (add-on) | Adds infographic parsing to a run | +$4 |
| Word bounding boxes (add-on) | A box and confidence score for every word | +$2 |
| Table cell bounding boxes (add-on) | Per-cell table bounding boxes. Free: it turns on word bounding boxes, which bill | +$0 |
| List bounding boxes (add-on) | Per-item list bounding boxes. Free: it turns on word bounding boxes, which bill | +$0 |

A "+" rate is an add-on on top of the processor it runs with. Details and plans: https://www.datalab.to/pricing.md

## Benchmarks

- Overall (OmniParseBench): Datalab accurate 94.1, Datalab balanced 92.6, Reducto 90.8.
- Tables (OmniParseBench · table tag): Datalab accurate 95.6, Extend 94.7, LlamaParse 93.7.
- Multi-column (OmniParseBench · multi_column tag): Datalab accurate 94.9, Datalab balanced 92.7, LlamaParse 91.4.
- olmOCR-bench (olmOCR-bench · Chandra 2.5): Chandra 2.5 86.9, Chandra OCR 2 85.8, LightOnOCR-3-4B 85.4.

Full tables with every parser, test counts, and sources: https://www.datalab.to/benchmarks.md

## Deployment and scale

- Managed cloud: capacity for 1B+ pages in a day, SOC 2 Type II, opt-in zero data retention, optional BAA and EU residency.
- VPC deployment on AWS, GCP, or Azure.
- On-prem and air-gapped, with full model weights.

## Open source

Marker, Surya, Chandra, and Indra: 76k+ combined GitHub stars. https://github.com/datalab-to

## Latest posts

- [Every flag needs evidence: how ReviewerZero AI traces integrity issues back to the page](https://www.datalab.to/blog/reviewerzero-datalab-case-study.md) (2026-11-13)
- [Chandra 2.7: better tables and languages](https://www.datalab.to/blog/chandra-2-7.md) (2026-11-12)
- [Everyone is #1: how to evaluate OCR for your own pipeline](https://www.datalab.to/blog/how-to-evaluate-ocr.md) (2026-11-12)
- [Fast mode now costs $1 per 1,000 pages, and reads tables better](https://www.datalab.to/blog/surya-2-6-fast-mode.md) (2026-11-11)
- [Evals: find the right mode for your documents](https://www.datalab.to/blog/evals.md) (2026-11-10)

Every blog post has a markdown version at its URL plus ".md".
