SnowPro Specialty: Gen AI Question 192
Single answerData augmentationA retail company is building a support chatbot in Snowflake to answer questions about thousands of product manuals, warranty policies, and troubleshooting guides stored as PDFs. Initial testing shows the model often gives incomplete answers because relevant details are spread across adjacent sections of the same document. The team wants to improve retrieval quality without retraining the model and while keeping the solution maintainable as documents are updated weekly. Which approach is MOST appropriate?
- A
Use data augmentation by chunking the documents with overlap, storing the chunks in a searchable knowledge base, and retrieving the most relevant chunks at inference time to provide grounded context to the model.
- B
Fine-tune the foundation model weekly on the full set of PDFs so the model memorizes the manuals and no longer needs retrieval.
- C
Increase the model temperature so it explores more possible answers and is more likely to include the missing details from nearby sections.
- D
Convert each PDF into a single large text row and pass the full document into the prompt for every user question, regardless of document size.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use data augmentation through retrieval over well-prepared document chunks. In GenAI systems, data augmentation commonly means enriching the model's prompt with relevant enterprise data at inference time rather than changing the model weights. For document-heavy use cases, best practices include parsing documents, chunking them into semantically useful segments, adding overlap when adjacent passages may both matter, and retrieving the most relevant chunks for a user query. This pattern is commonly referred to as retrieval-augmented generation (RAG). It is especially appropriate when source content changes frequently, because updates can be handled by reprocessing and reindexing documents instead of retraining a model. This aligns with generally recommended GenAI architecture guidance in Snowflake-oriented workflows: use retrieval and grounding for current enterprise knowledge, and reserve fine-tuning for cases where behavior or style must be adapted beyond what prompting and retrieval can accomplish.
- A. Correct.
Correct. This is the standard retrieval-augmented generation (RAG) style data augmentation pattern for enterprise document Q&A. Chunking with overlap helps preserve context when relevant information spans neighboring sections, while retrieval at inference time injects only the most relevant passages into the prompt. This improves answer grounding without retraining and is easier to maintain when source documents change frequently.
- B. Incorrect.
Incorrect. Fine-tuning is not the best fit here because the problem is primarily retrieval and grounding over frequently changing documents, not task specialization. Weekly retraining would be costly, operationally heavier, and may still not provide reliable citation-level grounding compared with retrieving current document chunks at inference time.
- C. Incorrect.
Incorrect. Temperature controls randomness and creativity in generation, not factual recall from external documents. Raising temperature would generally make outputs less deterministic and could increase hallucinations rather than improve completeness from adjacent source sections.
- D. Incorrect.
Incorrect. Passing entire documents for every query is inefficient and often exceeds practical context-window limits, especially with thousands of manuals. It also increases latency and cost. The misconception is that more raw text automatically leads to better answers; in practice, targeted retrieval of relevant chunks is the scalable approach.