COF-C03 Question 131
Single answerCortex SearchA retail company stores millions of product descriptions, support articles, and return-policy documents in Snowflake. The support team wants to build a chatbot that can answer natural-language questions such as "Can I return shoes after 45 days?" and retrieve the most relevant passages from this content with low operational overhead. The data engineering team wants a managed Snowflake feature that supports semantic retrieval directly over Snowflake data without exporting documents to an external vector database. Which solution best meets these requirements?
- A
Create a Cortex Search service on the document data and have the chatbot query that service for relevant results.
- B
Create a materialized view with ILIKE filters on the text columns and use it as the chatbot's retrieval layer.
- C
Use Snowpipe Streaming to ingest the documents, because streaming ingestion automatically provides semantic search over the loaded text.
- D
Create a dynamic table that tokenizes the text into keywords, because dynamic tables natively rank results by semantic similarity.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Search. In Snowflake, Cortex Search is intended for building managed search services over data already stored in Snowflake, making it well suited for chatbot and retrieval-augmented generation scenarios. This directly addresses the company's need for semantic retrieval without exporting data to an external vector database. By contrast, materialized views and dynamic tables are useful for optimization and transformation, but they do not provide semantic search capabilities. Snowpipe Streaming is strictly for ingestion. Snowflake documentation and product guidance position Cortex Search as the managed retrieval option for relevant document search over Snowflake data, especially when building AI-powered applications.
- A. Correct.
Correct. Cortex Search is a managed Snowflake capability designed for retrieval over Snowflake data, including semantic search use cases for applications such as chatbots and retrieval-augmented generation workflows. It allows teams to build a search service directly on Snowflake data instead of moving the content into an external search or vector system. This aligns with the requirement for natural-language retrieval with low operational overhead.
- B. Incorrect.
Incorrect. A materialized view with ILIKE-based filtering can support basic keyword matching, but it does not provide semantic retrieval. Users might choose this option because it seems simple and stays inside Snowflake, but it would not reliably return conceptually relevant passages when the question wording differs from the source text.
- C. Incorrect.
Incorrect. Snowpipe Streaming is an ingestion technology, not a semantic retrieval or search capability. It helps load data into Snowflake with low latency, but it does not create embeddings, rank text by meaning, or expose a managed search interface for chatbot retrieval.
- D. Incorrect.
Incorrect. Dynamic tables automate incremental transformation pipelines, but they are not a native semantic search engine. Tokenizing text into keywords may help with traditional text matching, but dynamic tables do not inherently provide semantic similarity ranking. This option reflects a common misconception that any text preprocessing feature can replace a purpose-built search service.