Hanzo AI Format - Universal .ai file format for AI artifacts

This crate provides a standard format for packaging, distributing, and sharing AI artifacts across the Hanzo/Zoo network. The .ai format supports:

• Model weights and biases
• Fine-tuning deltas (LoRA, QLoRA)
• Quantized models (GGUF, AWQ, GPTQ)
• Datasets and evaluation data
• Embeddings and vector stores
• Agent state and memory
• Configuration and hyperparameters

File Structure

artifact.ai (ZIP archive with .ai extension)
├── manifest.json          # Artifact metadata and manifest
├── data/                   # Primary artifact data
│   ├── weights/            # Model weights (safetensors, bin, etc.)
│   ├── config/             # Model configuration
│   └── tokenizer/          # Tokenizer files
├── delta/                  # Fine-tuning deltas (optional)
├── dataset/                # Training/eval datasets (optional)
├── embeddings/             # Pre-computed embeddings (optional)
├── state/                  # Agent state/memory (optional)
└── signatures/             # Cryptographic signatures

Storage Backends

The format supports multiple storage backends:

• Local filesystem
• HuggingFace Hub (fallback/mirror)
• P2P swarm (BitTorrent-style distribution)
• IPFS (content-addressed storage)

Example

use hanzo_ai_format::{AiArtifact, ArtifactType, Storage};

// Create a new artifact
let artifact = AiArtifact::builder()
    .name("my-model")
    .artifact_type(ArtifactType::Model)
    .add_weights("weights.safetensors", weights_data)
    .build()?;

// Save to .ai file
artifact.save("my-model.ai").await?;

// Upload to storage
let storage = Storage::new_with_hf_fallback();
storage.upload(&artifact).await?;