SnowPro Specialty: Gen AI Question 111
Single answerCortex SearchA retail company is building an internal support assistant on Snowflake to help agents find relevant troubleshooting articles and policy documents. The documents are stored in a Snowflake table with columns for document_id, title, body, product_line, region, and last_updated. The team wants to use Cortex Search so that agents can ask natural-language questions, but results must be restricted to the agent's region and should prioritize the most relevant passages rather than simple keyword matches. Which approach best meets these requirements?
- A
Create a Cortex Search Service on the documents table, index the text content, and apply a filter on the region attribute at query time so retrieval is limited to documents for that region.
- B
Create a standard Snowflake view filtered by region and use SQL LIKE queries against the body column, because Cortex Search does not support filtering on metadata.
- C
Store embeddings manually in a separate table and query them directly with SQL, because Cortex Search only supports exact-match search over VARCHAR columns.
- D
Use Cortex Complete directly on the full documents table without retrieval, because the model can infer the correct region and relevance from the prompt alone.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Search as the retrieval layer because it supports semantic search over Snowflake data and can incorporate structured filters such as region to narrow results. This is aligned with retrieval-augmented generation best practices: first retrieve relevant, authorized context, then use an LLM if needed for answer synthesis. Compared with SQL keyword matching, Cortex Search is better for natural-language questions because it focuses on relevance beyond exact token overlap. Compared with a custom embedding pipeline, it reduces implementation and maintenance burden by providing a managed search capability in Snowflake. And compared with prompting an LLM directly over no retrieved context, it provides stronger grounding, fresher enterprise data access, and better control over scoped retrieval. These patterns align with Snowflake guidance for using Cortex Search for semantic retrieval and metadata-aware filtering in GenAI applications.
- A. Correct.
Correct. Cortex Search is designed for semantic retrieval over enterprise data in Snowflake, which is exactly what this scenario requires. A search service can be built over document content and can use structured attributes such as region for filtering at query time. This allows the application to enforce region-specific retrieval while still benefiting from semantic ranking, which is better suited than keyword matching for natural-language agent questions.
- B. Incorrect.
Incorrect. This option reflects a common misconception that metadata filtering must be handled outside Cortex Search. In practice, Cortex Search is intended to support semantic retrieval with filtering over structured metadata. A simple SQL LIKE approach would degrade search quality because it relies on literal text matching and would miss semantically relevant content that does not share the same keywords.
- C. Incorrect.
Incorrect. This distractor assumes Cortex Search is limited to exact matching and that teams must manually manage embeddings to achieve semantic search. Cortex Search abstracts retrieval infrastructure so customers do not need to build and maintain a custom embedding search pipeline for this use case. Manually implementing vector logic adds operational complexity and is not the best answer when managed Cortex Search fits the requirement.
- D. Incorrect.
Incorrect. Using Cortex Complete alone skips retrieval entirely and would force the model to reason over data it was not explicitly given at inference time. That creates risks around hallucination, stale knowledge, and inability to reliably enforce region-scoped access to source documents. Best practice for enterprise question-answering is to retrieve grounded, relevant documents first and then optionally pass them to an LLM.