問題文
A team stores embeddings for two million photographs and must find the closest matches to a text query. A full comparison against every stored vector meets the accuracy requirement but is too slow. What is the standard trade-off they should consider?
選択肢
- Reducing the number of stored photographs until the full comparison fits the time budget: an exact answer over a smaller collection is always preferable to an approximate answer over the whole collection when users expect the top result to be correct.
- Storing the photographs themselves in the index instead of the embeddings.
- An approximate nearest-neighbor index, which returns most of the true closest items much faster in exchange for occasionally missing one.
- Recomputing the embeddings with a smaller model at query time.