SnowPro Specialty: Gen AI Question 108
Single answerEMBED_TEXT_768A retail company stores 20 million product descriptions in Snowflake and wants to build a semantic search feature entirely in SQL. The team plans to generate embeddings once, store them in a table, and compare incoming customer queries against those stored vectors. They want a built-in function that produces a consistent vector size suitable for text similarity calculations without calling an external model endpoint. Which approach best meets this requirement?
- A
Use EMBED_TEXT_768 to generate a 768-dimension embedding for each product description and each incoming query, then compare vectors using Snowflake vector similarity functions.
- B
Use AI_COMPLETE to generate a summary of each product description and compare the summaries with standard string matching functions such as LIKE.
- C
Use PARSE_DOCUMENT to convert product descriptions into JSON and compare the JSON structures to determine semantic similarity.
- D
Use EMBED_TEXT_768 only for incoming queries and compare those query embeddings directly to the original VARCHAR product descriptions.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use EMBED_TEXT_768 for both the stored product descriptions and the incoming user queries. EMBED_TEXT_768 is designed to create fixed-length text embeddings that can be stored and used for semantic similarity and retrieval scenarios inside Snowflake. This aligns with common retrieval patterns: precompute embeddings for relatively static content, compute query embeddings at request time, and compare vectors using Snowflake's vector capabilities. The other choices confuse embeddings with text generation, document parsing, or plain string matching. Snowflake documentation and best practices for vector search emphasize using embeddings for both corpus items and queries in the same vector space before applying similarity calculations.
- A. Correct.
Correct. EMBED_TEXT_768 is the built-in Snowflake function intended to convert text into a fixed-size 768-dimension embedding vector. This is the right choice when the goal is semantic similarity or semantic search using vectors stored in Snowflake. In a practical implementation, the team would generate embeddings for the corpus once, store them, embed each user query at runtime, and then use vector similarity calculations to rank the closest matches.
- B. Incorrect.
Incorrect. AI_COMPLETE is for text generation or transformation tasks, not for producing vector embeddings for similarity search. Summaries compared with LIKE or other lexical string operations do not provide semantic vector search behavior and would perform poorly for synonymy, paraphrasing, and broader semantic matching.
- C. Incorrect.
Incorrect. PARSE_DOCUMENT is used for document parsing and extraction workflows, not for creating semantic embeddings from text. JSON structure comparison does not measure semantic similarity between natural-language descriptions and is not an appropriate substitute for embeddings.
- D. Incorrect.
Incorrect. Vector similarity requires both sides of the comparison to be represented as vectors in the same embedding space. Embedding only the incoming query while leaving product descriptions as raw text would not allow valid vector-based similarity calculations.