Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
RAG & Document Chat

Xberg Review: Open-Source Document Intelligence for LLM and RAG Pipelines

·8 min readBy Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Xberg is an MIT-licensed document-intelligence engine written in Rust that extracts text, tables, metadata and structured data from about 107 file formats, web pages, audio and source code. You use it through one of 15 language bindings, a CLI, a REST server or an MCP server; it is a developer library, not a desktop app.

Key Takeaways

  • The library is free under the MIT licence; the maker also sells Pro, Enterprise and a managed Platform, none with public prices
  • It is the rebrand of Kreuzberg, rebuilt under a new v1 line, so older Kreuzberg write-ups describe the same engine
  • It needs a terminal or code: there is no desktop installer, and the install is one command per language
  • Speed and accuracy rankings come from the maker's own benchmarks and were not reproduced by PromptQuorum
  • The docs do not say whether it works fully offline, whether models download at first run, or whether it sends telemetry

📍 In One Sentence

Xberg turns documents, web pages, audio and code into LLM-ready Markdown and structured data, and its open-source core runs on your own CPU.

💬 In Plain Terms

Xberg reads almost any file you give it and hands back clean text, tables and metadata that a language model or a search index can use. You install it like a normal programming library or run it as a small server. If you want it to pull structured fields out with a language model, you can point it at a model running on your own computer.

Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

Xberg (xberg.io, documentation at docs.xberg.io, source at github.com/xberg-io/xberg) is a document-intelligence engine that converts files, web pages, audio and code into Markdown and other structured output. It belongs in a local-AI directory because the extraction, OCR and layout steps run on your own CPU and its LLM-backed extraction can talk to Ollama, LM Studio or vLLM instead of a cloud model. This review covers what Xberg does, what runs locally and what does not, how it is installed, and who it fits.

Xberg Review: Open-Source Document Intelligence for LLM and RAG Pipelines

What Is Xberg?

Xberg is a document-intelligence engine: you point it at a file, URL, archive or source tree, and it returns clean text, tables, metadata and structured data in a format an LLM or search index can use directly. Its README calls it "the fast, precise document-intelligence engine — for every language."

  • Product type: a Rust core with 15 language bindings, a 14-command CLI, a REST API (xberg serve), an MCP server, a Docker image and a Helm chart
  • Repository: github.com/xberg-io/xberg, 9,000+ stars, created in January 2025, verified via the GitHub API
  • Lineage: the README states that Xberg "is the next iteration of Kreuzberg" and the same engine "rebuilt and rebranded under a fresh v1 line"; the old kreuzberg-dev/kreuzberg address now redirects here, and a separate kreuzberg-lts repository still exists
  • Licence: MIT, per the repository's LICENSE file
  • Maker: the site shows the brand "Xberg" and the GitHub organisation xberg-io; its imprint gives a Berlin, Germany address and names one person as responsible for the website, but no legal entity name or registration number

What Can You Do With Xberg?

Xberg bundles the steps of a document pipeline into one engine. Some capabilities are optional feature flags that the prebuilt packages and the Docker image include by default.

  • Extract — read PDFs, Office files, images, HTML, email, e-books, scientific papers and structured data, with MIME detection and limits for large inputs
  • OCR — Tesseract, PaddleOCR, Candle or vision-language-model backends, with fallback chains and confidence scores
  • Layout and tables — ML layout and table-structure models rebuild reading order and cell grids as Markdown
  • Transcribe — speech-to-text for MP3, M4A, WAV, WebM and MP4 audio through Whisper ONNX models (feature flag)
  • Fetch URLs — extract one web page or crawl and follow links through the crawlberg engine (feature flag)
  • Unpack archives — recurse through nested zip, tar, gz and 7z files, guarded by zip-bomb, compression-ratio and nesting-depth limits
  • Code intelligence — functions, classes, imports and symbols from 371 programming languages, with syntax-aware chunking for RAG
  • Embed and enrich — local ONNX or provider-hosted embeddings, reranking, entity and keyword extraction, summarisation, translation and redaction
  • Structured extraction — schema-driven JSON from a local or hosted LLM
  • Output — plain text, Markdown, Djot, HTML, a JSON tree or Docling DocTags

What Runs Locally and What Does Not

Extraction, OCR, layout and ONNX embeddings run on your own CPU; the network is involved only where you choose a hosted provider, fetch a URL or use the maker's live demo. The docs describe the engine as CPU-first with no GPU required.

  • CPU: on x86_64, the ONNX-based features (PaddleOCR, layout detection, embeddings, reranking, transcription) need AVX/AVX2 and can crash without it; ARM is unaffected
  • Platforms: prebuilt binaries cover Linux (x86_64, aarch64), Windows and Apple Silicon; Intel Macs get prebuilt packages only for the CLI and the Python wheel
  • Browser build: the WebAssembly package runs in browsers, Deno, Cloudflare Workers and Node.js 22 or later, without the LLM and VLM features

Extraction, OCR, layout

Where it runs:
On your CPU, in-process or in your own server

Embeddings and reranking

Where it runs:
Local ONNX models, or a hosted provider you configure

Structured extraction

Where it runs:
Ollama, LM Studio or vLLM locally, or a hosted LLM

URL fetching and crawling

Where it runs:
Needs network access to the target site

Live demo on the docs site

Where it runs:
Uploads go through the maker's API

Not verified: the docs do not say whether models or OCR engines download at first run, whether the engine works fully offline, or whether it sends telemetry. Test those on an air-gapped copy before relying on them.

Xberg Licence and Commercial Editions

The library is free under the MIT licence for personal, internal and commercial use; the maker also offers paid editions without published prices. Details come from xberg.io and its privacy policy.

  • Xberg Pro: a self-hosted, single-container edition; Xberg Enterprise: a distributed Kubernetes edition with governance and support
  • Xberg Platform: a managed backend priced by usage or self-hosting; an invitation-only design-partner programme offers 24-month locked pricing
  • Website privacy: the policy, dated 21 May 2026, names "Xberg" as controller without naming a legal entity and says analytics run only with cookie consent
  • The policy makes no statement about the open-source library sending data to the maker

Who Should Use Xberg?

Xberg fits developers who need many document types converted into clean, structured text and would rather call one engine than stitch together separate PDF, OCR and layout libraries.

Use Xberg if you

  • Build RAG, search or agent pipelines that ingest PDFs, scans, Office files, e-mail or audio
  • Want extraction on your own hardware, in your language, as a library, server or MCP tool
  • Need LLM-based field extraction through a model server on your own hardware

Look elsewhere if you

  • Want a point-and-click app to chat with your documents
  • Need documented offline and telemetry guarantees today
  • Need a vendor with a named legal entity and published enterprise pricing

Competitors and Alternatives

The directory lists no other document-extraction engine, so the nearest same-category reviews are tools that fetch or consume content rather than convert files. Xberg sits upstream of them: it produces clean text and structure, and they store, search or act on it.

Firecrawl

Best known for:
Turning websites into LLM-ready Markdown or JSON
Articles about Firecrawl (2)

Cognee

Best known for:
Self-hosted long-term memory for agents via a knowledge graph

txtai

Best known for:
Embedded vector and LLM database in one library
Articles about txtai (2)

Also mentioned in:

See the Local LLM Software Directory for the full catalog, including Xberg's own entry.

Compare in directory

Frequently Asked Questions

Can Xberg run as an MCP server?

The README lists an MCP server next to the CLI and REST API. This review did not run it, so check the docs for the exact start command.

Can I use Xberg in a browser?

Yes, through the WebAssembly package, which also targets Deno, Cloudflare Workers and Node.js 22 or later. It leaves out the LLM and VLM features.

Does it run on an Intel Mac?

Partly. Prebuilt packages exist only for the CLI and the Python wheel, and the docs call Intel support best-effort and slated for removal.

Should I upload confidential files to the live demo?

No. The demo sends documents through the maker's API, so use a local install for anything sensitive.

Which languages can call Xberg?

Python, TypeScript and Node.js, WebAssembly, Rust, Go, Java, Kotlin on Android, C#, Ruby, PHP, Elixir, Dart and Flutter, Swift, Zig and C through an FFI.

Has PromptQuorum tested Xberg?

No. This review is based on the official site, documentation, privacy policy, GitHub repository and release data, not on hands-on benchmarks.

Sources

← Back to Power Local LLM

MCP