SnowPro Specialty: Gen AI Question 259
Single answerMinimize tokensA retail analytics team is building a chatbot in Snowflake that uses Cortex AI functions to answer questions about weekly sales trends. During testing, they notice high cost and latency because each request sends a long prompt containing repeated instructions, full table schemas, and several example interactions. The team wants to reduce token usage without materially hurting answer quality. Which approach is the BEST fit for this goal?
- A
Move stable instructions and repeated context into a concise system prompt, remove unnecessary few-shot examples, and retrieve only the most relevant business context for each question
- B
Increase the model size so the prompt can be processed more efficiently and keep the full schema and all examples in every request
- C
Convert all numeric fields in the prompt to text descriptions so the model sees more natural language and can compress the prompt internally
- D
Duplicate key instructions in both the system and user prompts to make sure the model does not miss them, even if this increases prompt length
Show answer and explanation
Correct answer: A
Explanation
The best answer is to reduce unnecessary prompt content and send only the information needed for the current request. In practice, minimizing tokens means using concise instructions, avoiding repeated text, trimming or eliminating unnecessary few-shot examples, and using retrieval to supply only relevant context rather than entire schemas or documents. These techniques lower cost and often improve latency while maintaining answer quality. This aligns with prompt-engineering and RAG best practices used with Snowflake Cortex AI functions and LLM applications more generally: keep prompts focused, avoid redundant context, and retrieve only the minimum relevant grounding data for each query.
- A. Correct.
Correct. Token minimization is primarily achieved by shortening prompts and sending only the context required for the current task. Consolidating stable guidance into a concise system prompt avoids repeating instructions, removing unnecessary examples reduces token overhead, and retrieving only relevant context instead of sending full schemas or large background text is a standard prompt-optimization and RAG best practice.
- B. Incorrect.
Incorrect. A larger model does not reduce the number of input tokens sent or billed. Keeping full schemas and all examples in every request directly increases token usage and often latency. This reflects a common misconception that model size solves prompt inefficiency.
- C. Incorrect.
Incorrect. Rewriting structured numeric data as longer natural-language descriptions typically increases token count rather than reducing it. The model does not 'compress' prompts automatically in a way that lowers billed input tokens. This option confuses model understanding with token efficiency.
- D. Incorrect.
Incorrect. Duplicating instructions across prompt sections usually increases token usage without proportional benefit. Good prompt design favors a single clear set of instructions, typically in the system prompt, rather than repeating the same content multiple times.