A Semantic Cache That Forgets Properly

Support queues ask the same question in a hundred phrasings. An exact-match cache never fires; a semantic one turns most of that traffic into a vector lookup instead of a generation. Which is a lot of money, and one subtle correctness question: what does "the same question" mean, and what do you throw away when the cache is full?

Write semantic_cache_hits(vectors, threshold=0.9, capacity=3).

Process vectors in order, as a stream of query embeddings. For each one, return the index of the cached query that served it, or -1 if it missed.

• A query hits if its cosine similarity to some cached entry is >= threshold. If several qualify, the most similar wins.
• A hit refreshes that entry's recency.
• A miss inserts the new query.
• Over capacity, evict the least recently used entry.

Return the list of results, one per input query.