Mistral Introduces New OCR API That Can Convert PDF Paperwork Into AI-Prepared Format

Mistral launched the Mistral Optical Character Recognition (OCR) software programming interface (API) on Thursday. The factitious intelligence (AI) mannequin is able to analysing and processing PDF paperwork and changing it into an AI-ready textual content format resembling Markdown or uncooked textual content file. The software is able to extracting information from PDFs to make them digestible for AI fashions. The Paris-based AI agency claimed that the Mistral OCR API will enable builders to construct AI purposes for PDF recordsdata in addition to enable them to create datasets to coach new AI fashions.

Mistral OCR API Launched

PDF paperwork pose a novel problem for AI fashions. The content material on this file format can’t be accessed by massive language fashions (LLMs) utilizing conventional Retrieval-Augmented Era (RAG) strategies as the info can’t be processed by them. For instance, if you happen to ask an AI software to scan by PDF paperwork in your laptop computer to discover a piece of data, it would battle to take action.

Which means builders constructing AI purposes shall be restricted in providing PDF-analysis functionality. Whereas Google’s NotebookLM, Adobe’s AI assistant, and a number of other different instruments use specialised OCR instruments to beat this problem, builders within the open-source group wouldn’t have entry to a high-efficiency software.

Mistral OCR API solves this problem by permitting builders to extract PDF information into an AI-ready format. The corporate claims in a newsroom put up that the software can perceive separate components in paperwork, together with media, textual content, tables, and equations with excessive accuracy. As soon as analysed, it may possibly extract and current the knowledge within the Markdown or a uncooked textual content file format.

AI fashions can then use this extracted textual content as enter and RAG techniques can simply entry them and reply queries about them. “Mistral OCR excels in understanding advanced doc components, together with interleaved imagery, mathematical expressions, tables, and superior layouts resembling LaTeX formatting. The mannequin permits deeper understanding of wealthy paperwork resembling scientific papers with charts, graphs, equations and figures,” the put up acknowledged.

The corporate claimed that the Mistral OCR can course of as much as 2,000 pages per minute on a single node. The API additionally lets builders use the doc as a immediate, and chain outputs to construct operate calling instruments and AI brokers.

Primarily based on inside testing, the Mistral OCR outperformed fashions resembling Google Doc AI, Azure OCR, and GPT-4o model 2024-11-20 for “text-only” paperwork. It additionally outperformed Google and Azure in multilingual capabilities.

These interested by attempting out the aptitude of the mannequin can go to Mistral’s Le Chat platform. The API will be accessed from la Plateforme.