Scope-Aware Cache Boundaries for Private Knowledge Retrieval

The article argues that a privacy-focused personal knowledge manager must treat its cache as part of the retrieval contract, since caching too early can leak context after permissions change while caching too late repeats costly searches. It recommends staged retrieval with explicit collections, bounded queries, and traceable sources, caching only candidate IDs behind a scope-aware key before reranking and citation. Using a media archive as the example, it illustrates how a cache key should capture every input that alters candidates, such as a hashed query fingerprint, collection revision, permission scope, and filters, without storing raw private text. The author also presents a three-boundary design (ingestion, bounded retrieval, reranking) with trace IDs and metrics that isolate which stage causes poor answers. The piece mentions Infrai as one possible vector-stage option for teams wanting a single REST contract.