xberg-io /xberg
The fast, precise document-intelligence engine for every language: text, tables and metadata from 100+ formats, with OCR, transcription and code intelligence.
★ 9,590⑂ 617👁 32⑃ 1,498● 1,631▤ 127
+5
日增长
9,590
stars
617
forks
32
watchers
1,498
prs
1,631
issues
127
releases
README
Xberg
The fast, precise document-intelligence engine — for every language.
Point Xberg at anything — a PDF, a scanned image, a spreadsheet, an audio file, a URL, a whole archive, or a source tree — and get back clean text, tables, metadata, and structured data. One engine handles format detection, reading, OCR, and extraction, so you never stitch a pipeline together from a dozen libraries.
107 formats · 141 file extensions · 371 code languages · 15 language bindings · 6 output formats · OCR · transcription · embeddings
The fastest, most precise open-source document and PDF-to-Markdown engine — see the benchmarks.
Install · What you get · Capabilities · CLI · Docs
Xberg is the next iteration of Kreuzberg. Same document-intelligence engine, rebuilt and rebranded under a fresh v1 line.
---
What you get
Point Xberg at anything — a PDF, a spreadsheet, a scanned image, an audio file, a URL, an archive, a source tree — and get back clean, structured content you can use right away. One core does the format detection, reading, and extraction, so you don't assemble a pipeline yourself. Call it from Rust, Python, Node.js, Go, Java, C, Ruby, PHP, Elixir, Dart, Swift, Zig, WASM, Kotlin, or C FFI, and run it as a library, CLI tool, REST API, or MCP server.
| Capability | What you get |
|---|---|
| 107 document formats | PDFs, Office, images, HTML, email, e-books, scientific publications, and structured data across 141 file extensions, with intelligent MIME detection and bounded extraction controls. |
| URLs & the web | Point Xberg at an http(s) URL — it fetches and extracts a single document, or crawls and follows links (Auto / Document / Crawl modes via the crawlberg engine). Requires the url-ingestion feature. |
| Audio & video transcription | Speech-to-text from MP3, M4A, WAV, WebM, and MP4 tracks via Whisper ONNX (tiny → large-v3). Requires the transcription feature. |
| Archives, traversed | List and recursively extract nested .zip, .tar, .gz, .7z — documents inside documents — guarded by zip-bomb, compression-ratio, and nesting-depth limits. |
| OCR on demand | Tesseract, PaddleOCR, Candle, or VLM backends — fallback chains, confidence scores, language auto-detection, extensible via plugins. |
| Layout & tables | ML layout models (PP-DocLayout-V3, RT-DETR) and table structure (TATR, SLANet) reconstruct reading order and cell grids for clean Markdown. |
| Code intelligence | Functions, classes, imports, symbols, docstrings from 371 programming languages. Syntax-aware chunking for RAG pipelines. |
| Embeddings & search | Local (ONNX) or provider-hosted embeddings (165 providers via liter-llm), sparse and late-interaction, cross-encoder reranking. |Top issues
#2332fix(tables): preserve HTML headers, spans and table boundaries16 comments · 2026-10-30#1567chore(lint): pay down the size and complexity baseline14 comments · 2026-09-05#2502fix(ocr): tall-scan strip recovery covers only a narrow set of scans12 comments · 2026-11-05#2331fix(ocr): preserve math and scanned text; separate headers and footers7 comments · 2026-10-30#2503fix(ocr): table rows still split across lines after the exact-row claim5 comments · 2026-11-05#2335feat(onnx): add CPU classification and captioning backends3 comments · 2026-10-30#2334perf(onnx): qualify CPU classification and captioning models3 comments · 2026-10-30#2336feat(models): publish pinned ONNX artifacts and SDK inventory2 comments · 2026-10-30
Discussions
#1872Bounded owner-run PDF extraction: adapter approach and research trial1 comments · 2026-09-26#1274Option to merge image descriptions into content for single-field consumers (RAG pipelines)0 comments · 2026-07-19#1200semantic versioning and maintenance of v43 comments · 2026-07-05#1184Kreuzberg v4.9.9 vs Xberg + patch plan for OpenSSL CVE2 comments · 2026-06-30#1119lilbee: a single-executable local search engine you can talk to. It runs the models, crawls the web into your library, and reads 90+ document formats with Kreuzberg0 comments · 2026-06-14#952Release frequency question2 comments · 2026-05-13#842License Question1 comments · 2026-04-29#737Preserve links in markdown for images?0 comments · 2026-04-17
Latest release
v1.4.0 v1.4.0 2026-11-13Highlights - Faster PDF extraction when only a few glyphs need OCR, with a full-resolution retry when needed. - Tesseract engine reuse for automatic script retries reduces repeated model loading. - Better preservation of native PDF text and exact page byte ranges when OCR falls back. - An experimental PP-DocLayout-V3 option for ONNX Runtime builds and explicit LaTeX annotations for Markdown, HTML, and Djot. Read the full v1.4.0 changelog.