Mistral Introduces New OCR API That Can Convert PDF Documents Into AI-Ready Format

Mistral introduced the Mistral Optical Character Recognition (OCR) application programming interface (API) on Thursday. The artificial intelligence (AI) model is capable of analysing and processing PDF documents and converting it into an AI-ready text format such as Markdown or raw text file. The tool is capable of extracting data from PDFs to make them digestible for AI models. The Paris-based AI firm claimed that the Mistral OCR API will allow developers to build AI applications for PDF files as well as allow them to create datasets to train new AI models.

Mistral OCR API Introduced

PDF documents pose a unique challenge for AI models. The content in this file format cannot be accessed by large language models (LLMs) using traditional Retrieval-Augmented Generation (RAG) techniques as the data cannot be processed by them. For example, if you ask an AI application to scan through PDF documents in your laptop to find a piece of information, it might struggle to do so.

This means that developers building AI applications will be limited in offering PDF-analysis capability. While Google’s NotebookLM, Adobe’s AI assistant, and several other tools use specialised OCR tools to overcome this challenge, developers in the open-source community do not have access to a high-efficiency tool.

Mistral OCR API solves this challenge by allowing developers to extract PDF data into an AI-ready format. The company claims in a newsroom post that the tool can understand separate elements in documents, including media, text, tables, and equations with high accuracy. Once analysed, it can extract and present the information in the Markdown or a raw text file format.

AI models can then use this extracted text as input and RAG systems can easily access them and answer queries about them. “Mistral OCR excels in understanding complex document elements, including interleaved imagery, mathematical expressions, tables, and advanced layouts such as LaTeX formatting. The model enables deeper understanding of rich documents such as scientific papers with charts, graphs, equations and figures,” the post stated.

The company claimed that the Mistral OCR can process up to 2,000 pages per minute on a single node. The API also lets developers use the document as a prompt, and chain outputs to build function calling tools and AI agents.

Based on internal testing, the Mistral OCR outperformed models such as Google Document AI, Azure OCR, and GPT-4o version 2024-11-20 for “text-only” documents. It also outperformed Google and Azure in multilingual capabilities.

Those interested in trying out the capability of the model can go to Mistral’s Le Chat platform. The API can be accessed from la Plateforme.