LibrePower Docling

LibrePower - Unlocking Power Systems through open source. Unmatched RAS and TCO. Minimal footprint 🌍

Document AI for AIX and Linux on Power - powered by IBM's Docling framework.

Early Release: These packages are provided as-is for testing and evaluation. While we use them in production, bugs may exist.

---

Join the Community

LibrePower is more than AIX—we're building open source support across the entire IBM Power ecosystem: AIX, IBM i, and Linux on Power (ppc64le).

Subscribe to our newsletter for releases, technical articles, and community updates.

librepower.org — Launching February 2026

---

Overview

LibrePower Docling brings enterprise document intelligence to IBM Power systems. Process PDFs, extract text and tables, and build document AI applications - all on-premise with your existing infrastructure.

Features

• 📄 PDF Processing: Extract text, tables, and document structure
• 🤖 AI-Powered: HuggingFace transformers and tokenizers for NLP
• 🖥️ Multi-Platform: Works on AIX (IBM Power) and Ubuntu Linux ppc64le
• 🔒 On-Premise: Data stays in-house - no cloud dependency
• ⚡ Power Optimized: Compiled for POWER9+ (runs on POWER9, POWER10, POWER11)

Quick Start

# Clone the repository
git clone https://gitlab.com/librepower/docling.git
cd docling

# Run the installer (auto-detects platform)
./install.sh

# Verify installation
./install.sh --verify

# Run the demo
./install.sh --demo

Supported Platforms

Platform	Status	Notes
AIX 7.3+	✅ Supported	POWER9, POWER10, POWER11
Ubuntu 22.04+	✅ Supported	ppc64le (Little Endian)

Installation Options

./install.sh              # Full installation (auto-detects processor)
./install.sh --power9     # Force POWER9 wheels (compatible with P9/P10/P11)
./install.sh --power10    # Use POWER10 wheels (MMA enabled, P10/P11 only)
./install.sh --deps-only  # System dependencies only
./install.sh --python-only # Python packages only
./install.sh --wheels-only # Install pre-built wheels only
./install.sh --verify     # Verify installation
./install.sh --demo       # Run demonstration
./install.sh --platform   # Show detected platform
./install.sh --help       # Show help

The installer auto-detects your processor and selects the appropriate optimized wheels.

Usage

Basic PDF Processing

from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("document.pdf")

# Get markdown output
print(result.document.export_to_markdown())

Low-Level API (docling-parse)

from docling_parse.pdf_parsers import pdf_parser_v2

parser = pdf_parser_v2()
parser.load_document("doc", "document.pdf")
result = parser.parse_pdf_from_key_on_page("doc", 0)

# Extract text from cells
cells = result['pages'][0]['original']['cells']['data']
text = ''.join([cell[12] for cell in cells])
print(text)

Enterprise RAG System

See examples/rag_demo.py for a complete Retrieval-Augmented Generation system.

from examples.rag_demo import PowerDocIntelligence

system = PowerDocIntelligence()
system.ingest_directory("/path/to/pdfs")
results = system.search("What are the AI capabilities?")

Pre-built Wheels

Pre-compiled wheels for AIX ppc64 are included in wheels/aix/.

POWER9 Wheels (Default - Compatible with P9/P10/P11)

Located in wheels/aix/:

Package	Version	Size	Compiler Flags
numpy	2.4.1	7.9 MB	-mcpu=power9 -mtune=power9 -mvsx
scipy	1.17.0	33 MB	-mcpu=power9 -mtune=power9
scikit-learn	1.8.0	10 MB	-mcpu=power9
hnswlib	0.8.0	3.1 MB	-mcpu=power9 -O3
annoy	1.17.3	102 KB	-mcpu=power9
libopenblas	0.3.28	18 MB	POWER9 BLAS library

POWER10 Wheels (Optimized for P10/P11)

Located in wheels/aix/power10/:

Package	Version	Size	Compiler Flags
numpy	2.4.1	7.9 MB	-mcpu=power10 -mtune=power10 -O3
hnswlib	0.8.0	438 KB	-mcpu=power10 -mtune=power10 -O3
scipy	1.17.0	33 MB	P9 fallback*
scikit-learn	1.8.0	10 MB	P9 fallback*
annoy	1.17.3	102 KB	P9 fallback*

*Some packages use POWER9 wheels as fallback (compilation issues with -mcpu=power10)

Note: OpenBLAS POWER9 was compiled specifically for this project - it did not exist for AIX before.

Processor Auto-Detection

The installer automatically detects your processor:

./install.sh              # Auto-detects P9/P10/P11
./install.sh --power9     # Force POWER9 wheels
./install.sh --power10    # Force POWER10 wheels (P10/P11 only)

If POWER10 wheels fail to install, the installer automatically falls back to POWER9 wheels.

POWER9/10/11 Compatibility

Processor	POWER9 Wheels	POWER10 Wheels
POWER9	✅ Native	❌ Won't run
POWER10	✅ Compatible	✅ Native + MMA potential
POWER11	✅ Compatible	✅ Native + MMA potential

POWER9 wheels provide:

• Full binary compatibility across P9/P10/P11
• VSX vector extensions for SIMD acceleration
• Major performance improvement over POWER7 binaries

POWER10 wheels additionally enable:

• MMA (Matrix Math Accelerator) potential for future NumPy/SciPy optimizations
• Native instruction scheduling for P10/P11 microarchitecture

Benchmarks

Performance comparison on POWER9 (same physical hardware, same LPAR config):

Test	AIX 7.3	Ubuntu 22.04 ppc64le	Notes
PDF Parsing	6.04 pages/s	4.56 pages/s	AIX 1.32x faster
Multiprocessing (32 workers)	19.41 files/s	4.34 files/s	AIX 4.47x faster

Important: These benchmarks measure end-to-end workload performance, not raw CPU compute. Differences are likely due to OS scheduling, I/O subsystem (JFS2 vs ext4), and Python multiprocessing behavior - not CPU performance. Both use custom POWER9-optimized OpenBLAS.

📖 Full benchmark methodology - Detailed test environment, reproduction steps, and caveats.

Project Structure

librepower-docling/
├── install.sh           # Universal installer (auto-detects P9/P10/P11)
├── wheels/
│   └── aix/             # POWER9 wheels (default)
│       ├── numpy-2.4.1-cp312-cp312-aix_ppc64.whl
│       ├── scipy-1.17.0-cp312-cp312-aix_ppc64.whl
│       ├── ...
│       └── power10/     # POWER10 optimized wheels
│           ├── numpy-2.4.1-cp312-cp312-aix_ppc64.whl
│           └── hnswlib-0.8.0-cp312-cp312-aix_ppc64.whl
├── lib/
│   ├── aix/
│   │   ├── install.sh   # AIX installer with P9/P10 support
│   │   ├── patches/     # XCOFF binary patches
│   │   └── shims/       # Compatibility shims
│   └── ubuntu/
│       └── install.sh   # Ubuntu installer
├── examples/
│   ├── quick_start.py   # Basic usage
│   └── rag_demo.py      # Enterprise RAG system
└── docs/
    └── SETUP_GUIDE.md

AIX-Specific Notes

XCOFF Binary Patching

The IBM Rust SDK generates XCOFF binaries with a loader relocation bug. The installer automatically patches the tokenizers binary.

Compatibility Shims

• pypdfium2: Uses Ghostscript for PDF rendering (PDFium not available on AIX)
• rtree: Pure Python spatial index (libspatialindex not available)
• numpy.f2py: Stub module (f2py not needed at runtime)

Known Issues

• docling-parse may segfault during cleanup (use os._exit(0) workaround)
• sklearn has import issues on AIX (investigating)
• Some packages may have version metadata mismatches

Contributing

Contributions welcome! Please see CONTRIBUTING.md for guidelines.

External Resources

IBM Python Ecosystem for POWER

For Ubuntu ppc64le, IBM provides optimized wheels at github.com/ppc64le/pyeco:

pip install --extra-index-url https://wheels.developerfirst.ibm.com/ppc64le/linux numpy scipy

LibrePower complements this with AIX ppc64 (Big Endian) support.

License

GPL-3.0 - See LICENSE file

Credits

Developed by SIXE as part of the LibrePower initiative.

Technologies

• IBM Docling - Document AI framework
• IBM pyeco - Optimized Python packages for POWER
• HuggingFace - Tokenizers and Transformers
• OpenBLAS - BLAS library

---

LibrePower - Unlocking Power Systems through open source

librepower.org | Newsletter | GitLab