Xberg — Document Intelligence Engine

Xberg is the successor to Kreuzberg — a Rust document intelligence engine that extracts text, tables, structured data, and metadata from 101 file formats across 115 file extensions. Written by Na'aman Hirschfeld, it's available as a library, CLI (13 commands), REST API (xberg serve), and MCP server (9 tools), with Docker and Helm deployment support.

Architecture

• Front-controller pattern: MIME detection → Extractor registry (priority-based) → Format-specific extractor → InternalDocument
• Plugin system: 8 plugin types (DocumentExtractor, OcrBackend, PostProcessor, Validator, EmbeddingBackend, RerankerBackend, TokenizerBackend, Renderer), all stored as Arc<dyn Trait> in global registries
• Engine seam injection: EngineBuilder pattern with 6 replaceable seams — CacheBackend, ProgressSink, StructuredPolicy, PresetResolver, LlmClient, ModelProvider — each defaulting to a built-in noop or core implementation
• Feature flag architecture: Over 100 Cargo features controlling format support, OCR backends, ML capabilities, and deployment modes; aggregate targets like full, server, wasm-target, android-target, windows-target, mobile

Extraction Pipeline

1. MIME detection → Extractor selection by priority → Format-specific extraction → InternalDocument
2. Image OCR processing (async) → Captioning prepass (runs BEFORE main derivation to feed captions into InternalDocument)
3. Derive ExtractedDocument → Image re-encode → Base64 pass → Early post-processors → Language detection
4. Provisional chunking (so Middle/Late processors see chunks) → Middle post-processors → Late post-processors → Token reduction → Validators → NFC normalization
5. Discard diverged internal/formatted content (#286, #331 fixes) → LLM structured extraction → Output format → Final chunking (after all content mutations — #213 fix for chunk offset correctness)

OCR Backends

• Tesseract (native C FFI + WASM), PaddleOCR (ONNX + tract), Sceptre (EasyOCR Gen2), Candle (pure-Rust transformer VLM), VLM (GPT-4V/Claude/Gemini)
• Inference backends: ONNX Runtime (default native), Sonos tract (pure-Rust, WASM/Android safe), Candle (pure-Rust transformer)

Language Bindings

15 language bindings via the "alef" polyglot binding generator (alef.toml): Rust, Python, Node.js, Go, Java, C#, Ruby, PHP, Elixir, Dart, Swift, Zig, WASM, Kotlin, C FFI. The alef system scans Rust source files and emits idiomatic bindings per language.

Code Intelligence

371 programming languages via tree-sitter, with dynamic loading on native platforms.

Embeddings & ML

Dense (ONNX/static), Sparse (SPLADE), Late-interaction (ColBERT/MaxSim), with reranking. NER via ONNX, Candle, or LLM. Layout detection, captioning, summarization, translation, classification, QR code detection, auto-rotate, document orientation.

Deployment Modes

• Library (embed in Rust projects)
• CLI (xberg) with 13 subcommands
• REST API server (xberg serve)
• MCP server (9 tools for agent integration)
• Docker & Helm
• WASM browser (via xberg-wasm)

Key Numbers

• v1.1.0, Rust edition 2024, MSRV 1.91
• 20 workspace member crates
• ~1340-line feature flag configuration in the core crate alone
• Chunking runs TWICE in the pipeline (provisional early, final late after all mutations)
• Content-hash caching, per-file timeouts, parallel batch processing

Licensing

Dual-licensed: MIT OR Apache-2.0 (unlike Kreuzberg's Elastic License 2.0 restriction).

Source: raw/xberg Last updated: 2026-08-07