Staged caching design for a privacy-focused personal knowledge manager

The article argues that a privacy-focused personal knowledge manager should use a staged retrieval architecture with a bounded cache, explicit collections, and cited source context. It recommends caching normalized retrieval results rather than private generated answers, paired with a freshness rule so stale pages are not presented as current. Each answer should be traceable through a retrieval contract that specifies the retrieval unit, metadata filters, and an acceptable freshness window. The author describes using a plain REST vector interface from a Go worker and keeping cache keys, encryption at rest, TTLs, and revision-based invalidation as explicit design choices. Ingestion and querying are treated as separate observable stages with request IDs and failure classes recorded for each.