Web Workers
Goal:
fast, reactive UI
run datatable sort in background
Embeddings Models
all‑MiniLM‑L6‑v2
Pros:
- Fastest
- Small/lean Cons:
- 78% top 5 retrieval acc, less than others
bge‑small‑en‑v1.5
Pros:
- Fast, not fastest
- quant 8bit
- 85% top 5 retrieval acc, high
- Prefix pattern → more acc results Cons:
- heavier
Choose this for accuracy over speed + prefix helps to naively navigate through sparse data. Since I don’t have the tech knowledge to provide the most relevant/cleanest RAG data for process knowledge.
Apple Neural Network
TODO: look into how to optimize for this
RAG
- Installer footprint – every extra MB ships to every user.
- First‑run latency – how long it takes to hydrate the local DB.
- Cross‑platform friction – anything that breaks on macOS notarisation or Windows AV is a headache.
- Runtime query performance – once the app is running you still need fast ANN look‑ups.