Document extraction that speaks more languages than most developers

What it does Kreuzberg extracts text, metadata, and structured data from more than 90 file formats—PDFs, Office documents, images, archives, and code—through a single Rust core. It exposes that engine through native bindings for more than a dozen languages, letting you call it from Python, TypeScript, Go, Java, Zig, Elixir, and others without leaving your runtime. The project ships as a library, CLI tool, REST API server, or MCP server.

The interesting bit Instead of treating document parsing as a language-specific chore, Kreuzberg bundles traditional extraction with code intelligence: it uses tree-sitter to pull functions, classes, and symbols from 300+ programming languages, then serializes results to a custom TOON wire format that claims 30–50% fewer tokens than JSON for LLM pipelines. It also passes HTML directly into Markdown conversion rather than forcing lossy intermediate round-trips.

Key highlights

• Native bindings for more than a dozen languages and runtimes, with precompiled binaries for x86_64 and aarch64 on Linux and macOS, plus Windows for most targets.
• OCR via Tesseract (including a WASM build for browsers), PaddleOCR, EasyOCR (Python only), and optional VLM OCR through 143 vision-model providers including local engines like Ollama.
• LLM integrations for structured JSON extraction with schema constraints and embeddings through 143 providers, powered by the sister liter-llm project.
• Plugin architecture for custom OCR backends, validators, post-processors, and renderers.
• GFM-quality Markdown output via Comrak, with proper table nodes, fenced code blocks, and bracket escaping.

Caveats

• WASM builds exclude ORT-dependent features such as PaddleOCR, layout detection, embeddings, and auto-rotate, and cannot run server modes.
• Precompiled macOS binaries are Apple Silicon only; Intel Macs are absent from the platform matrix.
• The Elastic-2.0 license is source-available but not OSI-approved, which may complicate commercial redistribution.

Verdict Teams building RAG pipelines or document-processing microservices in polyglot environments should look here; developers who only need occasional PDF text extraction in a single language will find it overbuilt.