Xberg
lastmod 2026-08-13created 2026-08-131 edits
A multilingual document intelligence framework
About
- Feed it anything (PDFs/images/spreadsheets/audio/URLs/archives/source trees, etc.) and it returns clean text/tables/metadata/structured data
- A single core handles everything from format detection to OCR and extraction, so there is no need to build a pipeline out of multiple libraries
- Supports 101 formats / 371 code languages / 15 language bindings / 6 output formats; can also extract text from Korean .hwp/.hwpx files
- URL/web collection: given an http(s) URL, extracts a single document or crawls by following links (Auto/Document/Crawl modes, using the crawlberg engine)
- Audio/video transcription: converts speech→text from MP3, M4A, WAV, WebM, MP4 tracks with Whisper ONNX (tiny→large-v3)
- Recursive archive extraction: extracts even nested documents inside .zip, .tar, .gz, .7z
- OCR backends: choice of Tesseract, PaddleOCR, Candle, VLM (GPT-4/Claude/Gemini Vision, etc.), with fallback chains/confidence scores/automatic language detection
- Layout/table recognition: restores reading order and cell grids with ML layout models (PP-DocLayout-V3, RT-DETR) and table structure models (TATR, SLANet) → clean Markdown
- Code intelligence: extracts functions/classes/imports/symbols/docstrings from 371 languages with tree-sitter, supports syntax-aware chunking for RAG
- Embedding/search: local (ONNX) or hosted embeddings from 165 providers, sparse/late-interaction, cross-encoder reranking
- Enrichment: NER, keyword extraction (YAKE/RAKE), summarization, translation, PII masking (redaction), page classification, QR detection, language detection, token savings (TOON, 30–50% savings compared to JSON)
- Structured extraction: generates schema-based JSON from documents directly, without prompt engineering, using local (Ollama, LM Studio, vLLM) or hosted LLMs
- Can be deployed as a library, CLI (12 commands), REST API (xberg serve), MCP server, Docker, Helm
- MIT license