Kreuzberg

Kreuzberg is a document intelligence platform designed to extract structured, machine-readable content from a wide range of document formats for integration into AI agents, pipelines, and applications. It addresses the need for rapid, scalable document processing in AI engineering workflows, offering features such as layout detection, optical character recognition (OCR), table extraction, semantic chunking, and code intelligence within a single API call.

The platform supports over 97 file formats, including PDFs (both native and scanned), images (JPG, PNG), Microsoft Office files (DOCX, PPTX, XLSX), HTML, web content, plain text, and code files across 306 programming languages. Users can interact with Kreuzberg via API, SDK, CLI, or Docker, making it adaptable to various development environments. The service delivers structured JSON output and offers webhook delivery for asynchronous workflows, enabling seamless integration into embeddings pipelines or retrieval-augmented generation (RAG) frameworks.

Kreuzberg incorporates a high-performance Rust core that enables document processing in milliseconds, facilitating batch processing at scale and supporting the indexing of millions of documents efficiently. Its features include built-in OCR for scanned documents, semantic chunking for code, markdown, and plain text, token reduction, keyword extraction, and metadata enrichment. The platform also provides code intelligence by extracting functions, classes, imports, and symbols from code files, preparing them for semantic chunking and RAG pipelines. Embedding generation is available via Rust-native fastembed-rs with multiple presets and extensibility to custom models, and the system supports vision language models (VLM) as an OCR backend, with access to 146 large language model providers, including local models without requiring API key configuration.

Kreuzberg is available as a cloud-based service on a pay-as-you-go pricing model, with no seat requirements or monthly minimums. The first 10,000 pages are processed free, after which a per-page fee applies. For organizations with high-volume needs, discounted rates are available upon consultation. A self-hosted option is also offered, allowing users to run the platform on their own infrastructure with the same capabilities as the cloud version, ensuring that data never leaves the user's environment. Documents are processed in memory and deleted immediately after extraction, with no storage or indexing, and the service does not use customer data for model training or improvement.