Phase 05 — Local-First AI
This phase turns mechanisms into product capabilities. Local-first means the application, model, data, storage, update, and fallback paths remain understandable when connectivity or device resources change; it does not mean “put a chatbot in a tab.”
Branches
- Local-First, Offline & Multimodal AI — Service Workers, model storage, resumable downloads, offline lifecycle, image/audio pipelines, memory pressure, privacy, updates, and rollback.
- Browser LLMs, Embeddings & RAG — local documents, chunking, embeddings, persistent vector search, evidence, context budgets, and optional small-model generation.
Workbench milestones
v5 — Multimodal local-first: deliver offline-capable image inference and small-model speech-to-text with verified model installation, visible progress, cancellation, and device-aware fallbacks.
v6 — Local semantic search: import documents, create deterministic chunks and local embeddings, persist vectors in IndexedDB, and return ranked evidence. Retrieval-only is the required first product; local generation is optional and later.
What this phase establishes
- An install/update/verify/activate/rollback model lifecycle.
- Explicit storage, quota, eviction, privacy, and offline state behavior.
- Multimodal pipelines that preserve preprocessing and golden-output contracts.
- Retrieval quality measured independently from any generator.
- Capability-based refusal or fallback for devices that cannot run an optional model.
Exit criteria
- The installed feature survives a full offline reload after model acquisition.
- Interrupted downloads, corrupted shards, quota failures, updates, and rollback are tested.
- Semantic search reproduces ranked evidence from a versioned corpus and golden query set.
- Optional generation can be removed without breaking retrieval and reports its own resource costs.
Continue with Phase 06: Beyond the Browser.