Cloud models are convenient. Local models are quieter infrastructure — slower to set up, easier to trust with private archives once they work.
After building a full self-hosted knowledge base, the lesson sharpened: own the data path (notes, vectors, capture, backups). Generation can stay on fast APIs for daily use if cost and privacy tradeoffs are explicit. Local embeddings are usually enough to make retrieval feel like “your notes answering you.”
Purity — “everything must run offline” — is a different product. Most entrepreneurs need a system that survives Tuesday morning, not a GPU shrine. Document what is local, what is routed, and how you’d bring more on-prem when the budget or risk profile changes.
The mature pattern lives in the Self-Hosted Knowledge Base project and its Workbench notes.