Look up a cached LLM response by exact (model, prompt, params) match, like an HTTP cache checking a request against its cache key. Call this BEFORE issuing an expensive prompt to a model. Returns hit:false on a miss or an expired entry — in that case, run the prompt yourself and store the result with cache_set. On a hit, freshness mirrors HTTP semantics: fresh means use it as-is; stale means it is past its TTL but within its stale-while-revalidate window, so you may still use it but consider refreshing it.

Store a prompt/response pair in the cache. Call this AFTER getting a fresh response from a model that cache_get did not have. New entries are automatically linked in a similarity graph to existing entries with related content, so cache_related and cache_query can surface them later even without an exact key match. Pass derived_from with parent entry ids if this result was built from other cached entries — invalidating a parent with cascade can then clean up anything derived from it.

Semantic search across all cached entries, independent of exact key matching. Use this when you suspect something related was already computed even though the prompt wording differs — the knowledge-graph equivalent of a cache lookup by meaning instead of by exact key.

Traverse the cache graph outward from one entry to find connected entries — ones auto-linked for content similarity ("similar" edges) or explicitly declared as built on top of it ("derived-from" edges). Use this to pull in everything already known that connects to a given cached result.

Delete a cache entry, e.g. because the underlying source it was based on changed. With cascade:true, also deletes every entry that declared this one as a derived_from parent, and recursively theirs — dependency-graph invalidation instead of a manual hunt for stale copies.

Cache analytics: entry/edge counts, exact-match hit rate, queryHits/queryMisses for semantic recall, and tokensServed — the token count handed back from cache. Note that tokensServed equals tokens saved only for entries that stand in for work which would otherwise be regenerated; serving a cached file read costs the same tokens as re-reading the file, so it saves nothing.