Mistral OCR 4 returns a document map, not a text dump

Most OCR APIs hand you a wall of text and wish you luck with chunking.

Mistral OCR 4 , released June 23, 2026, does something more useful for production pipelines: it returns a map of the document. Every block gets a bounding box, a type label, a confidence score, and reading-order placement.

On human evaluations across 600+ real-world documents, Mistral reports a 72% blind-test win rate against leading OCR systems. I treat vendor win rates as marketing until I run my own contracts and intake forms through the API. The structural output is the part I would pay for regardless of the headline number.

Segmentation beats transcription for RAG

RAG quality often dies in the parser, not the embedding model.

When a PDF becomes one long string:

• Tables collapse into nonsense tokens
• Headers lose hierarchy
• Footnotes pollute retrieval chunks
• Agents cite the wrong paragraph because nothing is localized

Mistral OCR 4 returns typed blocks in reading order:

Bounding boxes localize each region on the page. Confidence scores flag blocks that need human review before they enter a compliance workflow.

That is the difference between "we OCR'd it" and "we can ground an agent answer to a rectangle on page 4."

API shape I would actually wire

Endpoint: POST https://api.mistral.ai/v1/ocr

import os
from mistralai import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

response = client.ocr.process(
    model="mistral-ocr-latest",
    document={
        "type": "document_url",
        "document_url": "https://example.com/contract.pdf",
    },
    include_blocks=True,
    confidence_scores_granularity="block",
)

for page in response.pages:
    for block in page.blocks:
        print(block.type, block.content[:80], block.top_left_x, block.top_left_y)

Mistral OCR processor docs document the full block schema. Block extraction requires OCR 4 or newer; older models accept include_blocks but return empty arrays.

For structured extraction without a second LLM pass, use document_annotation_format with a JSON schema. Invoice fields, clause types, and form keys can come back typed in one call.

Pricing and deployment options

Availability includes Mistral Studio, Amazon SageMaker, and Microsoft Foundry.

For teams that cannot let documents leave the VPC, self-hosting is the enterprise checkbox. Mistral markets a single container deployment for OCR 4, which is the right framing for air-gapped legal and clinical workflows.

Where this fits next to Surya and Textract

I already run Surya OCR 2 locally when clients need Apache-friendly weights on a GPU they control. Surya wins on sub-3B parameter efficiency and olmOCR-bench scores at 650M scale.

Mistral OCR 4 wins when you want:

• Managed API with block geometry and confidence in one response
• 170-language coverage with gains on rare scripts
• DOC/PPT ingestion without a conversion step
• Agent workflows that highlight source regions in the UI

Do not assume one OCR tool wins every doc type. Run your worst PDFs: scanned leases, handwritten intake, multi-column SEC filings, and phone photos of whiteboards.

Production checklist

1. Chunk on block types, not fixed token windows. Titles and tables should not share embedding space with body text.
2. Route low-confidence blocks to human review queues before they enter retrieval indexes.
3. Store bounding boxes in your metadata even if the UI does not use them yet. Compliance teams will ask for highlight overlays eventually.
4. Batch overnight at $2/1k pages when latency allows.
5. Compare citation accuracy on agent answers before and after block-aware chunking. That is the metric that pays for the upgrade.

If you are rebuilding a document RAG pipeline and want help choosing between local OCR, Mistral blocks, and embedding strategy, book a free discovery call.