Retrieval caches: validate the model, corpus, and access dependencies

A document embedding depends on source content, preprocessing, and the embedding model artifact. A retrieval result additionally depends on query representation, index generation, retrieval configuration, filters, and principal scope. A packed context depends on selected source versions, order, rendering, tokenizer, and packing policy. A generated answer depends on those inputs plus the generation model, prompt, decoding configuration, and any tools or conversational state affecting the result. Store these dependencies explicitly rather than reusing one query-only key at every stage.

Canonicalize key fields carefully and retain enough structured metadata to detect incompatible reuse. A content hash can identify immutable text but does not imply matching permissions, model behavior, or source authority. A model alias that changes behind an unchanged name can invalidate reproducibility assumptions. Some dependencies can be represented by a conservative pipeline version or corpus generation, at the cost of more misses. Finer invalidation can improve reuse only when its dependency tracking is complete and its compatibility rules are tested.

A matching key is a candidate for reuse, not proof that current permission remains valid. Validate the artifact's source dependencies and required access policy before delivery to each consumer. A changed ACL can revoke access without changing document text, while a changed corpus can alter the top results even for a query whose previous result did not mention the updated document. Invalidation based only on old result membership can therefore miss newly relevant additions or ranking changes. Negative-result caches need updates too.

A TTL bounds age only under its actual enforcement rules; it does not establish correctness until expiration or immediate revocation. Event-based invalidation can be delayed or lost, so use an authoritative version or validation mechanism that meets the serving contract. Distinguish immutable historical snapshots, which may remain valid for an authorized historical request, from answers claiming to reflect the current corpus. HTTP cache directives in RFC 9111 have specific semantics; an application-level RAG cache still needs its own source and permission dependency design.

Test a content edit, new relevant document, source deletion, permission-only change, embedding-model replacement, prompt update, and generator change. Include caches populated by another principal and in-flight invalidation races. Verify that citation resolution and derived summaries follow the same snapshot and policy rules as answer serving. Report valid-hit rate separately from raw key matches, alongside stale-response and authorization failures. A lower hit rate after correct invalidation can be the intended outcome rather than a performance regression.

The cache lab demonstrates only deterministic LRU replacement. Reset, set C=2 and s=4, predict zero hits among sixteen initial requests, then Play or Step. Separately Reset, set C=4 and s=4, predict twelve initial hits, then Play or Step. Both cycle over four addresses, and only the larger capacity holds them all. The instrument has no corpus generations, ACLs, model versions, TTLs, or invalidation events. Use it for capacity-versus-working-set intuition while testing dependency validity in the actual pipeline.