Question 140
A production Snowflake Gen AI workflow must handle the following: A claims reviewer wants a reasoned customer response grounded in verified case facts. Which solution best fits the problem without ignoring governance or evaluation?
Explanation
Correct Answer
“Invoke an appropriate Cortex COMPLETE generation workflow with task-specific context and evaluate its free-form answer against requirements” is supported by this requirement: A claims reviewer wants a reasoned customer response grounded in verified case facts. General text generation serves drafting and synthesis tasks that are not reducible to a fixed label or embedding. Verify it using reference-scored task accuracy, groundedness and response latency on varied representative prompts.
Incorrect Answers
“Use SENTIMENT to write the requested original narrative text” would not meet the requirement that a claims reviewer wants a reasoned customer response grounded in verified case facts. General text generation serves drafting and synthesis tasks that are not reducible to a fixed label or embedding. The more reliable approach is to invoke an appropriate Cortex COMPLETE generation workflow with task-specific context and evaluate its free-form answer against requirements.
“Use EMBED_TEXT alone and interpret its vector as a human-readable response” would not meet the requirement that a claims reviewer wants a reasoned customer response grounded in verified case facts. General text generation serves drafting and synthesis tasks that are not reducible to a fixed label or embedding. The more reliable approach is to invoke an appropriate Cortex COMPLETE generation workflow with task-specific context and evaluate its free-form answer against requirements.
“Use a SELECT COUNT(*) query to generate a prose recommendation without text generation” would not meet the requirement that a claims reviewer wants a reasoned customer response grounded in verified case facts. General text generation serves drafting and synthesis tasks that are not reducible to a fixed label or embedding. The more reliable approach is to invoke an appropriate Cortex COMPLETE generation workflow with task-specific context and evaluate its free-form answer against requirements.