Question 80
A developer is testing a Snowflake Gen AI workflow with this requirement: A support engine ranks incident embeddings to retrieve the closest historic resolution. Which solution has the clearest match to the described task and constraints?
Explanation
Correct Answer
“Rank same-space vectors by VECTOR_COSINE_SIMILARITY with the direction and normalization assumptions checked” is supported by this requirement: A support engine ranks incident embeddings to retrieve the closest historic resolution. Cosine similarity is higher for more aligned vectors, given valid comparable embeddings. Verify it using top-k relevance and ranking consistency on a human-labeled similarity test set.
Incorrect Answers
“Order cosine-similarity results ascending when the intended match is most similar” would not meet the requirement that a support engine ranks incident embeddings to retrieve the closest historic resolution. Cosine similarity is higher for more aligned vectors, given valid comparable embeddings. The more reliable approach is to rank same-space vectors by VECTOR_COSINE_SIMILARITY with the direction and normalization assumptions checked.
“Compare vectors of incompatible dimensionality and treat the errors as relevance rankings” would not meet the requirement that a support engine ranks incident embeddings to retrieve the closest historic resolution. Cosine similarity is higher for more aligned vectors, given valid comparable embeddings. The more reliable approach is to rank same-space vectors by VECTOR_COSINE_SIMILARITY with the direction and normalization assumptions checked.
“Use an uncalibrated absolute difference in document text length as semantic similarity” would not meet the requirement that a support engine ranks incident embeddings to retrieve the closest historic resolution. Cosine similarity is higher for more aligned vectors, given valid comparable embeddings. The more reliable approach is to rank same-space vectors by VECTOR_COSINE_SIMILARITY with the direction and normalization assumptions checked.