SnowPro Specialty: Gen AI Question 194
Single answerData transformationsA team is building a retrieval-augmented generation (RAG) solution in Snowflake. They load product manuals into a table with one row per document, but search quality is poor because relevant sections are buried inside large documents. The team wants to improve retrieval accuracy before generating embeddings, while preserving document context and minimizing duplicate results. Which transformation approach is the most appropriate?
- A
Split each manual into smaller, semantically coherent chunks with slight overlap, store chunk metadata such as document ID and section order, and then generate embeddings for each chunk.
- B
Generate one embedding for the full manual and rely on the LLM at query time to identify the relevant section from the retrieved document.
- C
Duplicate each manual into several rows with the same full text but different manually assigned tags so vector search has more opportunities to match.
- D
Remove headings, page structure, and section boundaries before embedding so that only raw text remains and the model is not biased by formatting.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to chunk long source documents into smaller, semantically meaningful units before generating embeddings. In Gen AI workloads, data transformation is critical because embedding quality and retrieval quality depend heavily on how the source text is prepared. For RAG, chunking with limited overlap is a widely accepted best practice: it improves precision by matching queries to specific passages instead of entire documents, while overlap helps preserve meaning across boundaries. Storing metadata with each chunk is equally important for filtering, deduplication, ranking, and assembling the final context sent to the model. In Snowflake-based Gen AI pipelines, this transformation pattern aligns with practical guidance for preparing unstructured text for vector search and downstream generation: preserve useful structure, create retrieval-sized chunks, and maintain lineage back to the original document.
- A. Correct.
Correct. For RAG pipelines, a common best practice is to transform long documents into smaller chunks before embedding. Chunking improves retrieval granularity, and a small overlap helps avoid losing context at chunk boundaries. Keeping metadata such as document ID, title, and chunk or section order supports traceability, filtering, and reconstructing context after retrieval. This is the most practical transformation to improve search quality when the source table currently has one row per large document.
- B. Incorrect.
Incorrect. A single embedding for an entire long manual is often too coarse for accurate retrieval because the vector represents the full document rather than the specific passage relevant to a user query. This commonly leads to lower recall and poor ranking of precise content. Candidates may choose this because it seems simpler operationally, but it does not address the root problem of insufficient retrieval granularity.
- C. Incorrect.
Incorrect. Duplicating the same full text with different tags does not create better semantic coverage and can increase storage, maintenance effort, and duplicate retrieval results. This reflects a misconception that more rows automatically improve vector search. In practice, duplicate full-text rows usually add noise rather than meaningful retrieval improvements.
- D. Incorrect.
Incorrect. Removing structural signals such as headings and section boundaries usually harms retrieval rather than helping it. Titles, headings, and nearby context often improve the semantic meaning of a chunk. Candidates might pick this if they assume formatting is irrelevant, but document structure is often valuable during transformation for chunking and downstream retrieval.