ABBYY gives old-school OCR a job in the AI pipeline

ABBYY has packaged its FineReader OCR engine as a self-hosted tool for turning documents into structured text that AI systems can use. The company's new FineParser runs in a Docker container on a CPU, without requiring a GPU. It aims to preserve the layout of a document as it extracts its contents. It takes images of documents, in multiple languages, and turns them into structured, formatted text - so that they can be processed using, for instance, modern generative AI LLMs. This isn't its sole purpose: the company suggested it could help bring print text into a modern CMS, or for archiving, or as a stage in some form of production pipeline. ABBYY describes FineParser's approach as "deterministic" AI. It extracts text and document structure rather than generating a plausible rendition of them. Its output can then be passed to a generative AI system, whose responses are less predictable. The company also offers its own programmable machine-learning framework, NeoML, which is FOSS and available on GitHub. It also publishes an OCR SDK for companies wanting to embed the FineReader engine into their own products. FineParser itself isn't open source, although ABBYY maintains a GitHub repository for examples and community support. The self-hosted tool has a free tier allowing 1,000 pages per month for one year. ABBYY says its subscription tiers connect to a license server for validation; a fully offline deployment requires an Enterprise plan. Preserving structure means recognizing columns in reading order, headings, and tables - even those without borders - rather than returning a jumble of extracted text. ABBYY says FineParser also handles handwriting and more than 200 languages. Developers can submit documents through FineParser's REST API and receive output as plain text, JSON, or DocLang, a compact format intended for LLM input. FineParser's pitch is that, before an LLM can make use of a document, something has to read it properly. ABBYY is betting that its decades-old approach to OCR - running on a CPU, preserving the layout, and leaving the text-generating to other tools - still has a place in the AI pipeline. Sometimes the old-fashioned part is the useful part. (R)