Xberg is a Rust-core document-intelligence engine that extracts text, metadata, images, tables, and structured data from documents, images, audio, archives, URLs, and source trees, and it is built for developers and data teams who want one extraction core instead of a pipeline stitched together from many libraries.

What it is

Xberg is an open-source, MIT-licensed polyglot document-intelligence engine written in Rust, listed in the AI & Machine Learning / AI Development Platforms category. Its README describes it as the next iteration of Kreuzberg: the same document-intelligence engine, rebuilt and rebranded under a fresh v1 line. It ships fifteen language bindings and can be called from Rust, Python, Node.js, Go, Java, C#, Ruby, PHP, Elixir, Dart, Swift, Zig, WASM, Kotlin, or C FFI, and it runs as a library, a CLI tool, a REST API, or an MCP server. Project documentation lives at https://docs.xberg.io.

The concrete problem it solves is pipeline assembly. Format detection, reading, OCR, and extraction normally come from separate format-specific tools that a team has to wire together and maintain. Xberg handles all of it in one core: 107 document formats across 141 file extensions, 371 code languages, six output formats, plus OCR, transcription, and embeddings. Point it at a PDF, a scanned image, a spreadsheet, an audio file, a URL, an archive, or a source tree, and it returns clean text, tables, metadata, and structured data in one step.

Key capabilities

• Extraction across 107 document formats and 141 file extensions — PDFs, Office files, images, HTML, email, e-books, scientific publications, and structured data — with intelligent MIME detection and bounded extraction controls.
• OCR through Tesseract, PaddleOCR, Candle, or VLM backends, with fallback chains, confidence scores, language auto-detection, and plugin extensibility.
• Layout and table reconstruction using ML layout models (PP-DocLayout-V3, RT-DETR) and table structure models (TATR, SLANet) to rebuild reading order and cell grids for clean Markdown.
• Code intelligence for 371 programming languages, extracting functions, classes, imports, symbols, and docstrings, with syntax-aware chunking for RAG pipelines.
• Speech-to-text from MP3, M4A, WAV, WebM, and MP4 tracks via Whisper ONNX, spanning tiny through large-v3 models, gated behind the transcription feature.
• URL ingestion that fetches a single document or crawls and follows links in Auto, Document, or Crawl modes through the crawlberg engine, gated behind the url-ingestion feature.
• Schema-driven JSON extraction from any document via local Ollama, LM Studio, or vLLM, or hosted LLMs, alongside local ONNX embeddings or 165 provider-hosted options via liter-llm and cross-encoder reranking.

Who uses it and how

• RAG and retrieval teams use the syntax-aware chunking, local ONNX or provider-hosted embeddings, sparse and late-interaction retrieval, and cross-encoder reranking to build indexing pipelines without a separate chunker.
• Polyglot engineering organisations call the same Rust core from any of the fifteen bindings, avoiding per-language reimplementations of extraction logic.
• Platform teams self-host it through Docker or Helm, running a library, the 14-command CLI, an xberg serve REST API, or an MCP server, on CPU by default with no GPU required.
• Batch and archive workloads rely on content-hash caching, parallel batch execution, and per-file timeouts to process large document sets predictably.

Getting started

Install the packaged binding with pip install xberg for Python or npm install @xberg-io/xberg for Node.js and TypeScript, or deploy the Docker image or Helm chart. The engine can also be run as a CLI tool or as a REST API through xberg serve.

How it compares

Xberg does not compete with the extraction engines it wraps; it orchestrates them, drawing on Tesseract, PaddleOCR, Candle, Whisper ONNX, crawlberg, and liter-llm behind a single interface. It shares output ground with Docling by rendering DocTags among its six output formats, and it succeeds the Kreuzberg project as the same engine under a fresh v1 line.

When to use it — and when not to

Capabilities such as URL ingestion, transcription, reranking, and layout models sit behind Cargo feature flags, so a from-source build enables only what is selected, while prebuilt language packages and the Docker image bundle the common set. Teams that cannot run local ML models or want a hosted SaaS product with no operational surface should look elsewhere, since Xberg expects a self-managed runtime. The README excerpt also documents Python and Node.js installation in detail only, leaving other bindings to their per-package READMEs, so adopters of less common languages should check those before committing.