Skip to content

Latest commit

 

History

History
70 lines (51 loc) · 2.51 KB

File metadata and controls

70 lines (51 loc) · 2.51 KB

VerifyFetch Roadmap

Current: v1.1.0

Theme: Protect Every AI Model in Your Browser

New in v1.1.0:

  • Transformers.js integration (@verifyfetch/transformers package)
    • verifiedPipeline() - drop-in verified pipeline wrapper
    • enableVerification() - global verification for all Transformers.js downloads
    • preloadVerifiedModel() - pre-download and verify HF model files
    • Integrates via env.customFetch — no monkey-patching
  • CLI hash-model command - Generate manifests for any Hugging Face model
    • npx verifyfetch hash-model Xenova/distilbert-base-uncased-finetuned-sst-2-english
    • Auto-detects ML files (ONNX, tokenizer, config)
    • Supports --chunked for large model files
  • Pre-computed manifests (@verifyfetch/manifests package)
    • Ready-to-use hashes for popular Transformers.js and WebLLM models

Shipped in v1.0.0:

  • Streaming integrity verification (constant 2MB memory for any file size)
  • Chunked verification with fail-fast (detect corruption at chunk N, stop immediately)
  • Resumable downloads (persist to IndexedDB, resume on page reload)
  • Service Worker integration (automatic verification for all matching fetches)
  • Multi-CDN failover (sequential, race, fastest strategies)
  • CLI tools (sign, enforce, init)
  • Manifest system (v1 simple, v2 with chunked hashes)
  • WebLLM integration (@verifyfetch/webllm package)
    • VerifiedMLCEngine - drop-in replacement for MLCEngine with verification
    • preloadVerifiedModel - pre-download and verify before WebLLM loads
    • Model manifest format with chunked hashes for large shards
    • Addresses WebLLM issue #761

Test coverage: 338+ tests passing


Next: v1.2.0

Goal: Build plugins and broader ecosystem.

  • Next.js plugin for build-time manifest generation
  • Vite plugin for build-time manifest generation
  • ONNX Runtime Web integration
  • More pre-computed manifests for popular models

Considering

These are ideas, not commitments:

  • Ed25519 signatures - Verify publisher identity, not just file integrity
  • Worker thread hashing - Offload to Web Workers for large files
  • Delta verification - Only verify changed chunks
  • MediaPipe integration - Verified loading for MediaPipe models

Contributing

See CONTRIBUTING.md for development setup.

Good first issues:

  • More framework examples (SvelteKit, Remix)
  • Performance benchmarks
  • Documentation improvements
  • Pre-compute manifests for more popular models