LLM Compression Layer

Universal semantic compression layer for LLM inputs. Compresses memories, code, and context before API calls while preserving reasoning equivalence across Claude, GPT, and Gemini.

Quick Start

# Setup
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

Target Metrics

• Token reduction: 40-60%
• Task equivalence: >95%
• Compression latency: <50ms
• Cross-model transfer: >90%

Project Structure

compression-layer/
├── src/
│   ├── validation/     # Cross-model equivalence testing
│   ├── generation/     # Compression pair generation
│   ├── training/       # Fine-tuning pipeline
│   ├── inference/      # Production compressor service
│   └── utils/          # Tokenizers, caching, cost tracking
├── data/               # Corpora and generated datasets
├── models/             # Checkpoints, adapters
├── configs/            # YAML configs
├── scripts/            # Entry points
└── tests/

Environment Variables

Create a .env file with:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=...
HF_TOKEN=hf_...
TINKER_API_KEY=tk_...

License

MIT