One endpoint for every document you have
PDF, DOCX, PPTX, XLSX, EPUB, HTML, and image formats go through the same call. Multi-column spreads, digital files, and scans are all handled the same way.
PROCESSORS · CONVERT
The document parser that doesn't mangle your tables.
Most parsers split tables mid-row and turn scans into garbage. Convert keeps tables, equations, figures, and reading order intact — across PDFs, Office files, and scans, in 90+ languages.
Watch it read 250-year-old handwriting.
Convert can handle any document you throw at it. Every region it finds comes back with a bounding box, confidence score, and layout-aware, clean markdown.
1 TEXT2 H13 TEXT4 TEXT CONFIDENCE 0.0%
- 1 Text IN CONGRESS, JULY 4, 1776. 0.997
- 2 Section Header The unanimous Declaration of the thirteen united States of America. 0.994
- 3 Text · 5,432 chars When in the course of human events, it becomes necessary for one people to dissolve the political bands… 0.982
- 4 Text · Signatures Signatures of the delegates arranged in five columns. Column 1: John Hancock, Samuel Adams… 0.961
What Convert handles, out of the box.
The cases that break naive OCR
Colspan, rowspan, nested headers, and tables that run across pages stay intact. Printed, handwritten, and mixed-script equations come back in one pass.
Charts and figures come back as assets
Every figure, chart, and embedded image is extracted as a discrete asset and bound to the caption it belongs to, so nothing gets flattened into the surrounding text.
A confidence score on every parse
Convert returns a Parse Quality Score from 0 to 5 on each call, so you can drop low-confidence documents before they reach your index or your training set.
Default to markdown, or ask for the shape your pipeline wants.
Call Convert through the SDK or as a single HTTP request. One parameter sets the output format: markdown for LLM context, HTML for a CMS, JSON for block-level access, or chunks for an embedding pipeline.
from datalab_sdk import DatalabClient, ConvertOptions
client = DatalabClient()
# Default: markdown. No options needed for the common case.
result = client.convert("document.pdf")
print(result.markdown)
# Pre-chunked for embedding pipelines:
options = ConvertOptions(output_format="chunks", mode="balanced")
result = client.convert("document.pdf", options=options)
# result.chunks is a dict of pre-chunked spans — feed straight into
# your embedder + vector store.
embed_and_index(result.chunks)
# Parse Quality Score (0-5) — drop low-quality docs before indexing.
print(result.parse_quality_score)Better audit trails with word-level bboxes and confidence scores
Enable word bounding boxes and get the position and confidence score for each word. On the scan below, every word comes back above 0.98 confidence except one which gets flagged as low confidence because of an extra umlaut, making it easy to design better human review workflows.
Abhörchgerät · 0.015 garbled umlaut — route to reviewexpiration · 0.99998 clean read — auto-accept 0 words 0 flagged 0.0%
<p>
<span data-bbox="65 1372 219 1403" data-confidence="0.015">Abhörchgerät</span>
<span data-bbox="231 1372 332 1403" data-confidence="1">monitor,</span>
<span data-bbox="344 1372 409 1403" data-confidence="1">check</span>
</p>On non-English documents, Convert scores 80.4% to Gemini's 67.6%.
Convert runs on the same open models we publish — Marker, Surya, and Chandra. We benchmark them in the open and lead by the widest margin on the work most parsers quietly fail: documents that aren't in English. It reads 90+ languages, including vertical Chinese, Japanese, and Korean.
MULTILINGUAL · TOP 43 LANGUAGES · +12.8 VS GEMINI 2.5 FLASH
Rank Model Score
01 Datalab API 80.4%
02 Chandra 2 OSS 77.8%
03 Chandra 1 prior generation 69.4%
04 Gemini 2.5 Flash 67.6%
05 GPT-5 Mini 60.5%
TABLES · OLMOCR-BENCH · #1 AMONG OPEN OCR
Rank Model Score
01 Datalab API 90.7%
02 Chandra 2 OSS 89.9%
03 Chandra 1 prior generation 88.0%
Where Convert ships in production.
TRAINING CORPORA
Model training data
Clean training corpora for LLMs. Reading order, equations, citation graphs — preserved.
See the deep-dive →RAGRAG indexing
Chunk at section, table, and figure boundaries. Retrieval returns complete units of meaning.
See the deep-dive →PUBLISHINGWeb-ready PDFs
Convert long-form publications into responsive HTML or clean markdown. Headings intact.
See the deep-dive →Run Convert wherever your documents have to stay.
- Managed cloud
Start with an API key — nothing to host or operate.
Sign up → - EU data residency
Run in-region, with no egress to US infrastructure.
Sign up → - Your VPC
Runs inside your own AWS, GCP, or Azure account.
Talk to sales → - On-prem & air-gapped
Fully offline, the same model weights, dedicated support.
Talk to sales →
SOC 2 Type II · BAA available View our trust center →
START
Run Convert on your hardest PDF.
Free tier, no credit card. 100M+ pages a day, sustained — capacity is not the bottleneck.