Mistral OCR
Product
Mistral OCR is an optical character recognition API from Mistral AI that converts documents, including scans and tables, into structured Markdown; in the episode it runs in Azure AI Foundry.
Також обговорюють
У випуску на тему Mistral OCR обговорюють також: n8n, Chunking, Embeddings, Azure OpenAI Service, GDPR та Retrieval-augmented generation (RAG).
Description
Mistral OCR is an optical character recognition API from Mistral AI. It converts documents, including scans and tables, into structured Markdown so that their content can be searched, processed or passed on to other systems. In Mistral's documentation the capability appears under the name Document AI OCR processor, and the mistral-ocr-latest alias points to the newest OCR model. In the M365.FM episode on private retrieval augmented generation, it is used inside Azure AI Foundry.
The service extracts text and structured content from PDFs and images while preserving document structure and hierarchy. Supported inputs include images such as PNG, JPEG and AVIF, and documents such as PDF, PPTX and DOCX. A file can be passed as a public URL, as Base64 or as an uploaded file. The documentation lists support for more than 40 languages.
The result is returned as a JSON object that contains pages. Each page holds Markdown, images, tables, hyperlinks, header, footer, dimensions, confidence scores and blocks, and the response also includes model and usage information. Tables can be left inline in the Markdown or returned separately as Markdown or HTML tables through a table format parameter. Headers and footers can be extracted separately. Block extraction returns paragraph-level bounding boxes with structural labels such as title, list, table, equation, caption, code and signature, in reading order. Confidence scores can be requested at page, block or word level.
Some features depend on the model version. According to the documentation, table formatting and separate header or footer extraction require OCR version 2512 or newer, and block extraction requires OCR 4 or newer. For large volumes, Mistral recommends its Batch Inference service, and for structured output it points to Document Annotations. The documentation page itself does not mention Azure AI Foundry, so details of that hosting setup come only from the episode.
A typical situation is a knowledge base built from scanned or image-based documents. Text that exists only as a scan cannot be searched or retrieved by a language model until it has been recognized and converted. Mistral OCR addresses that step by producing Markdown that keeps headings, lists and tables, which is a more usable form for indexing than flat extracted text. Decisions around it usually concern the output format for tables, whether headers and footers should be kept or stripped, whether confidence scores are needed to flag uncertain passages, and whether batch processing is appropriate for the document volume.
The episode "How to Build Private RAG with SharePoint and n8n" (2026-06-06) places Mistral OCR in a private retrieval augmented generation setup. The episode states that tools like Mistral OCR help convert scanned documents into searchable data and thereby improve the knowledge base. It sits alongside correct SharePoint setup for document organization and retrieval, and workflow automation with n8n that connects SharePoint and manages data retrieval tasks. The episode also recommends regular testing and iteration so that the solution stays reliable and efficient, and notes that such a solution can be scaled by integrating multiple data sources and adapting workflows to an organization's needs.
Overview compiled with AI from M365.FM episodes and the facts on this page. You can improve it with the edit button below.
| Type | Product |
|---|
Community wiki
Short info above, community knowledge below: add practical tips, examples or corrections. New contributions are reviewed first; trusted members publish directly.
Experts and companies for this topic
Nobody has listed this topic yet. Add it to your profile.
Maintained by Mirko Peters (Microsoft MVP) · Last updated: