SnowPro Specialty: Gen AI Question 189
Single answerData enrichmentA retail company is building a customer-support copilot in Snowflake. Support agents want the copilot to answer questions using product manuals, warranty PDFs, and troubleshooting articles stored in Snowflake stages. The team notices that retrieval quality is poor because many chunks contain only partial sentences, missing product identifiers, or ambiguous references such as "this model" and "it." They want to improve answer relevance before generating embeddings, while keeping the pipeline inside Snowflake as much as possible. Which approach is the BEST fit for this data enrichment requirement?
- A
Use Cortex Search directly on the raw staged files and rely on semantic retrieval alone, because embeddings remove the need for preprocessing or metadata enrichment.
- B
Create a preprocessing pipeline that extracts document text, adds structured metadata such as product line and document type, and rewrites or summarizes chunks with enough context before generating embeddings and indexing them for retrieval.
- C
Increase the chunk size significantly so each chunk contains the entire source document, which avoids ambiguity and eliminates the need for metadata enrichment.
- D
Skip retrieval enrichment and instead fine-tune a model on all support documents, because model training is the preferred way to inject current enterprise knowledge into a copilot.
Show answer and explanation
Correct answer: B
Explanation
The best answer is Option 2 because the problem is fundamentally about data enrichment for retrieval quality. In a Snowflake GenAI pipeline, improving chunk content before embedding is often more effective than relying on raw semantic search alone. Practical enrichment steps include extracting clean text from documents, adding metadata fields that support filtering or ranking, and transforming ambiguous chunks into self-contained passages through contextual rewriting or summarization. These steps help embeddings capture the intended meaning and improve retrieval relevance in downstream RAG applications. This aligns with common Snowflake and GenAI best practices: use retrieval for enterprise knowledge grounding, enrich data before indexing, and avoid using fine-tuning as a replacement for current-document retrieval. Option 1 reflects the misconception that embeddings alone solve all document-quality issues. Option 3 reflects the common mistake of over-expanding chunk size, which can hurt precision and increase cost. Option 4 confuses model adaptation with knowledge grounding; current support content is usually better handled through enriched retrieval rather than model training.
- A. Incorrect.
Incorrect. Semantic retrieval helps, but raw document ingestion alone does not solve issues caused by poor chunk boundaries, missing context, or lack of metadata. In retrieval-augmented generation workflows, preprocessing and enrichment often improve relevance by making chunks more self-contained and easier to retrieve accurately.
- B. Correct.
Correct. This is the strongest data enrichment approach for the scenario. Extracting text, attaching metadata, and rewriting or summarizing chunks so they carry their own context can materially improve downstream embedding quality and retrieval precision. In Snowflake-based GenAI workflows, enriching content before indexing aligns with best practices for RAG pipelines, especially when source documents contain ambiguous references or inconsistent structure.
- C. Incorrect.
Incorrect. Very large chunks can reduce retrieval precision, increase token usage, and dilute the relevance of embeddings. While larger chunks may reduce pronoun ambiguity in some cases, embedding entire documents usually makes it harder to retrieve the most relevant passage and is not a substitute for targeted enrichment.
- D. Incorrect.
Incorrect. Fine-tuning is not the preferred solution for injecting frequently changing enterprise knowledge such as manuals and troubleshooting content. RAG with enriched data is generally better for maintaining current knowledge sources. Fine-tuning is more appropriate for changing model behavior or style, not for keeping a copilot up to date with evolving documentation.