## Problem Pinecone's query API has no server-side offset/skip: it always returns the same top-K nearest neighbors for a given vector and topK value, so passing $offset straight to Pinecone would silently be ignored and every "next page" request would come back byte-for-byte identical to the first. Instead, over-fetch topK = $limit + $offset in one request (matches are already ranked by score) and slice off the first $offset client-side, giving callers correct pagination-like behaviour despite the backend having no real concept of it. ## Proposed resolution Fetch topK + offset and slice