A team embedded 2 million product descriptions last year with the e5-base-v2 model and stored them in a VECTOR column. A new search feature embeds each query with snowflake-arctic-embed-m-v1.5, which returns vectors of the same dimension. Queries run without errors, but the top results are unrelated. What is the most likely cause?
- A.
Query and document vectors come from different models, so their positions in vector space are not comparable
- B.
The stored vectors are more than a year old, and embeddings expire unless the service refreshes them
- C.
The similarity call is placed in the SELECT clause, so it is evaluated after the LIMIT and misses rows
- D.
VECTORCOSINESIMILARITY ignores vector magnitude, so it cannot rank vectors produced by two different models
Show answer
Answer: A
Vectors from different embedding models are not comparable, even when they have the same dimension.
- A. Correct: similarity is only meaningful between vectors produced by the same embedding model; equal dimensions merely let the function run.
- B. Incorrect: stored vectors do not expire; they only need regenerating when the text changes or the model changes.
- C. Incorrect: Snowflake recommends the SELECT clause for the similarity call, with ORDER BY and LIMIT selecting the top matches.
- D. Incorrect: ignoring magnitude is how cosine similarity works for any vectors; the problem is that the two vector spaces differ.