SnowPro Specialty: Gen AI Question 97
Single answerEXTRACT_ANSWERA support analytics team stores product manuals and policy PDFs in a Snowflake table and uses a Cortex pipeline to answer employee questions. They want the system to return only a short answer span taken directly from the retrieved document text, not a generated summary or paraphrase. During testing, they notice that free-form LLM completions sometimes add unsupported wording. Which approach best addresses this requirement?
- A
Use EXTRACT_ANSWER on the retrieved document chunk together with the user’s question so the result is grounded in the source text
- B
Use COMPLETE with a higher temperature so the model can produce a more concise answer from the document
- C
Use EMBED_TEXT_768 to generate vectors and return the embedding with the highest similarity score as the final answer shown to users
- D
Use a larger text-splitting window so COMPLETE has more context and is less likely to paraphrase
Show answer and explanation
Correct answer: A
Explanation
The key requirement is extractive QA: returning a concise answer directly from source text. In Snowflake Cortex, EXTRACT_ANSWER is intended for this use case, where you already have relevant context and want a grounded answer span from that context. By contrast, COMPLETE is a text generation function and is appropriate for summarization, transformation, and open-ended answering, but it does not guarantee source-exact wording. Embedding functions such as EMBED_TEXT_768 support semantic retrieval, which is often an upstream step in RAG pipelines, but they do not produce final human-readable answers. A practical best practice is to use retrieval to find the most relevant chunk, then use EXTRACT_ANSWER when the application requires answers copied from the evidence rather than generated prose.
- A. Correct.
Correct. EXTRACT_ANSWER is designed for extractive question answering: given a passage and a question, it returns an answer span from the provided text rather than generating a new response. This is the best fit when the business requirement is to minimize hallucination and present wording grounded directly in the retrieved source.
- B. Incorrect.
Incorrect. COMPLETE is a generative function, so even with prompt tuning or different temperature settings, it can still paraphrase, summarize, or introduce unsupported wording. Lowering or raising temperature affects variability, not the fundamental behavior of extractive versus generative answering.
- C. Incorrect.
Incorrect. EMBED_TEXT_768 creates embeddings for semantic search and retrieval, not end-user answers. An embedding vector cannot be presented as a human-readable answer. This option confuses retrieval with answer extraction.
- D. Incorrect.
Incorrect. Increasing chunk size may improve retrieval context in some cases, but it does not ensure that the returned answer is an exact span from the source. COMPLETE remains a generative step and can still produce paraphrased or unsupported text.