SnowPro Specialty: Gen AI Question 112
Single answerCortex SearchA support engineering team stores product manuals and troubleshooting articles in a Snowflake table with columns DOCUMENT_ID, TITLE, BODY_TEXT, PRODUCT_LINE, LANGUAGE, and LAST_UPDATED. They want to build a Retrieval-Augmented Generation (RAG) assistant using Cortex Search so that users can ask questions in natural language and receive answers grounded only in documents for the selected product line and language. They also want search results to include the document title and last updated timestamp for display in the application. Which approach best meets these requirements with the least custom implementation?
- A
Create a Cortex Search service on the table using BODY_TEXT as the searchable text, define PRODUCT_LINE and LANGUAGE as filter attributes, and configure TITLE and LAST_UPDATED as attributes returned with results.
- B
Create a Cortex Search service on the table using TITLE as the searchable text, and rely on the LLM in the RAG application to infer the correct PRODUCT_LINE and LANGUAGE from the user prompt instead of applying structured filters.
- C
Create a dynamic table that concatenates TITLE, BODY_TEXT, PRODUCT_LINE, and LANGUAGE into one large text column, then create a Cortex Search service with no filterable attributes because semantic search will handle the restrictions automatically.
- D
Skip Cortex Search and use AI_COMPLETE directly against the BODY_TEXT column so the model can read the table contents and generate grounded answers for the requested product line and language.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Search as the retrieval layer with searchable content, structured filters, and returned metadata aligned to the application requirements. In this scenario, BODY_TEXT is the correct search corpus because users ask detailed natural-language questions whose answers are likely embedded in the main article text. PRODUCT_LINE and LANGUAGE are classic examples of structured dimensions that should be applied as filters during retrieval to ensure only relevant documents are considered. TITLE and LAST_UPDATED are metadata fields that are useful to return with search results so the client application can display citations, freshness indicators, or source references. This follows recommended RAG design patterns in Snowflake, where Cortex Search is used to index and retrieve relevant chunks or documents, and a generation model is used afterward to synthesize an answer grounded in those retrieved results. The incorrect options reflect common misconceptions: using only titles for search, expecting semantic search to replace explicit filters, or using a text-generation function as a substitute for retrieval.
- A. Correct.
Correct. This is the standard and most practical Cortex Search design for a RAG workflow. BODY_TEXT should be the primary searchable content because it contains the substantive manual and troubleshooting information users will query. PRODUCT_LINE and LANGUAGE should be modeled as structured filter attributes so the application can restrict retrieval to the requested subset instead of hoping the model infers constraints from prompt text. TITLE and LAST_UPDATED are useful metadata fields to return with results for display and citation in the application.
- B. Incorrect.
Incorrect. Searching only TITLE would significantly reduce retrieval quality because relevant troubleshooting content is typically in the document body, not just the title. Also, relying on the LLM to infer and enforce PRODUCT_LINE and LANGUAGE from the prompt is weaker than applying explicit structured filters in Cortex Search. This can lead to retrieval of irrelevant or cross-product documents, which undermines grounding.
- C. Incorrect.
Incorrect. Concatenating filter dimensions such as PRODUCT_LINE and LANGUAGE into the searchable text is a common but flawed workaround. Cortex Search supports structured filtering, which is more precise and operationally cleaner. Semantic search does not automatically enforce business constraints like product line or language unless those fields are explicitly used as filters. This design also makes metadata handling less clear and can reduce retrieval precision.
- D. Incorrect.
Incorrect. AI_COMPLETE is for text generation, not indexed retrieval over enterprise data. It does not replace a search index and cannot by itself perform efficient, filtered retrieval over document collections in the way Cortex Search is designed to support. For grounded RAG, the recommended pattern is to retrieve relevant documents first and then pass that context to a model for answer generation.