問題文
A retrieval system indexes photograph embeddings. The team decides to switch to a newer embedding model that produces vectors of a different dimension. What must happen to the existing index?
選択肢
- The index can be kept and the two models used in parallel, choosing per query, so each query is answered by whichever model produced the closer vector.
- Nothing, as long as both models were trained on similar data, so the two sets of vectors line up closely enough for the stored entries to keep working.
- Only the query side needs to change, since the index stores the vectors and the comparison happens at query time, so as long as the query vector is projected to the stored dimension the existing entries remain usable without any re-processing.
- Every stored photograph must be re-embedded with the new model, because vectors from two models do not share a space.