Document Processing
The European Patent Office needed a near-error-free way to transcribe millions of patent pages into ST36-compliant structured output, but standard OCR tools struggled with the formulas, tables, figures, multilingual text, and high-volume reliability requirements of patent processing.
Reported impact
The EPO reduced transcription lead times from days to minutes while maintaining less than 1% character recognition error at 400,000 pages per day, giving the office a scalable compliant workflow and freeing expert staff to focus on patent examination rather than manual transcription cleanup.
The European Patent Office worked with Mistral on a three-month proof of concept to automate patent transcription without losing the accuracy and standards compliance the grant process requires. At the core is a 1B OCR model fine-tuned on more than 150,000 PDFs representing 50,000 unique patents, so the system can read dense multilingual documents containing tables, formulas, figures, and scientific notation that defeat standard OCR tools. The model feeds a pipeline that converts outputs from Markdown to HTML and then into ST36 XML, producing structured patent data instead of plain extracted text. Mistral and EPO also delivered a scalable, fault-tolerant application ready for on-premises deployment, giving the office a reliable way to process very large document volumes while keeping the workflow aligned to operational and compliance requirements.
Patent PDFs containing text, tables, formulas, figures, and multilingual content are read by the 1B OCR model, converted through a Markdown-to-HTML-to-ST36 XML pipeline, and delivered as structured output for the patent grant process through a scalable application designed for reliable on-premises use.
"400,000 pages/day throughput"
"<1% character recognition error rate"
"Lead time from up to 5 days down to a few minutes"
"The core of the solution included a 1B OCR model, specifically fine-tuned on over 150,000 PDFs, representing 50,000 unique patents."
"The team developed a sophisticated pipeline that successfully converts complex, multifaceted patent documents from Markdown, to HTML, and finally into the required ST36 XML format."
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Mistral Sample May 27, 2026 HTX HTX needed AI that understood Singapore's public-safety context, could be trained on sensitive data inside a secure environment, and would remain under government control rather than depending on generic off-the-shelf… 123B-parameter natively multimodal and multilingual foundation model AI Forge · AI Studio · NGINE Mistral Sample May 25, 2026 Austrian Academy of Sciences Researchers at the Austrian Academy of Sciences needed a way to analyze and reconstruct huge volumes of fragmented Ancient Greek texts that manual transcription and interpretation could not process quickly or at… 1M+ unread Greek papyri worldwide, now accessible through AI Apollo 
Salesforce Sample Sep 3, 2026 California Department of Motor Vehicles The CA DMV needed to replace legacy, siloed licensing and service systems with an omnichannel platform that could handle identity services for tens of millions of residents while cutting wait times, manual paperwork,… 175+ field offices CRM Analytics · Experience Cloud · MuleSoft 