Decide deliberately: Embedding model — open small model on-device vs. API — must be on-device for privacy Vector store — sqlite-vec vs. a hosted DB — on-device for privacy LLM — hosted frontier vs. on-device small model — tradeoff above Eval — RAGAS/custom The innovator framing: pick the simplest thing that meets the privacy bar and quality bar, instrument it, and only add complexity (re-ranking, agentic retrieval, fine-tuning) when the eval numbers justify it. Show you'd ship a thin reliable slice and iterate on measured faithfulness.