SnowPro Specialty: Gen AI Question 191
Single answerData augmentationA retail company is building a support assistant in Snowflake using Cortex Search over product manuals, return policies, and troubleshooting guides. During testing, users ask vague questions such as "Can I send it back after opening?" and "Why is it blinking red?" The assistant often retrieves irrelevant chunks because the queries lack product context and use informal wording that does not match the source documents. The team wants to improve retrieval quality without changing the underlying manuals. Which approach is the MOST appropriate data augmentation strategy for this scenario?
- A
Preprocess incoming user questions to expand them with structured context such as product name, category, and likely synonyms before sending them to retrieval
- B
Lower the warehouse size used by the application so that fewer documents are scanned and the top result is more focused
- C
Convert all manuals into a single large document so the retrieval system has more complete context in each chunk
- D
Remove informal user wording from the query entirely and only pass exact keywords that appear in the manuals
Show answer and explanation
Correct answer: A
Explanation
The best answer is to augment the query before retrieval. In GenAI applications, data augmentation is not limited to generating more training data; it also includes enriching inputs used by retrieval pipelines so the system can better connect user intent with indexed knowledge. For a Snowflake implementation using Cortex Search, a common best practice is to improve retrieval quality through better chunking, metadata use, and query formulation. Here, the core issue is that user questions are ambiguous and phrased differently from the source material. Query augmentation with product context and synonyms directly addresses that problem while preserving the original manuals. By contrast, changing warehouse size is a compute optimization, not a relevance strategy; merging documents harms retrieval granularity; and reducing queries to exact keywords ignores the benefits of semantic retrieval. This aligns with retrieval best practices used in RAG systems and with Snowflake guidance around improving search relevance through high-quality indexed content, metadata, and effective query design.
- A. Correct.
Correct. This is a practical query augmentation pattern for retrieval-augmented generation workloads. If user questions are underspecified or phrased differently from the indexed content, enriching the query with known metadata and semantically related terms can improve recall and ranking. In this scenario, adding product identifiers, product category, and likely terminology variants helps Cortex Search match the relevant manual or policy content without altering the source documents.
- B. Incorrect.
Incorrect. Warehouse size affects compute resources and performance characteristics, not the semantic quality of retrieval. Reducing compute does not improve relevance and may degrade responsiveness. This option reflects a common misconception that infrastructure tuning can substitute for retrieval or prompt design improvements.
- C. Incorrect.
Incorrect. Combining all manuals into one large document generally makes retrieval worse, not better. Effective retrieval depends on well-structured, appropriately sized chunks so that relevant passages can be found precisely. A single large document reduces granularity and can increase irrelevant context returned to the model.
- D. Incorrect.
Incorrect. Stripping away informal wording and keeping only exact document keywords is usually harmful in this situation. Users often express issues in natural language that differs from official documentation. Retrieval systems benefit from preserving intent and, where useful, adding synonyms or context, rather than narrowing the query to brittle keyword matching.