Question 300
At fictional Larchwood City Services, a Snowflake data and AI engineering team is conducting a governed document-grounding pilot under documented decision rights and minimum-privilege security requirements. The separately composed technical incident is: A chat service concatenates full retrieved documents and conversation history until request sizes exceed model limits, causing frequent failures and inconsistent truncation. Assume only these documented facts and use safe, authorized production-change procedures; exact current Snowflake feature syntax and region or account availability require current review. Which repeatable assurance or control is most relevant to preventing this situation?
Explanation
Writer-proposed answer: D (UNVERIFIED)
Preliminary individual option explanations — NOT expert validated
Correct Answer
Add token telemetry, bounded retrieval packing and regression checks for large documents and long conversation histories. Context-window limits apply to complete assembled requests, so token budgeting and source prioritization are core design constraints.
Incorrect Answers
This response concerns Batch classification retry cost explosion, which is different from the stated failure. A chat service concatenates full retrieved documents and conversation history until request sizes exceed model limits, causing frequent failures and inconsistent truncation.
Answer extraction without source attribution is insufficient to demonstrate this result: Stress tests show accepted prompts remain under configured model limits and critical supporting evidence stays available for cited answers.
Unauthorized tenant content in retrieval is not the cause of the reported failure: A chat service concatenates full retrieved documents and conversation history until request sizes exceed model limits, causing frequent failures and inconsistent truncation.