フリー問題

問題文

A team is building similarity search over 50,000 knowledge articles. Some articles were embedded last year with one model and the rest this year with a different one. Users report that results are erratic. What explains this?

選択肢

1. Similarity scores are only valid for text under 512 characters, and the longer product descriptions fall outside the supported range and score at random.
2. The two models produce vectors of different lengths, so the comparison silently pads the shorter vectors with zeros and the padding dominates the resulting score.
3. Vectors produced by different embedding models are not comparable, so similarity computed across the two sets is meaningless.
4. The cosine similarity function requires normalized vectors, and one of the two model outputs is left unnormalized before the comparison.