SnowPro Specialty: Gen AI Question 243
Select 2Methods to reduce model hallucinations and biasA retail company built a customer-support assistant in Snowflake that uses a large language model to answer questions about return policies, warranty terms, and shipping rules. The team notices two issues in production: the assistant sometimes invents policy details that are not in the company documents, and its responses about premium vs. budget product lines occasionally use loaded language that could reinforce bias. The team wants to reduce both hallucinations and biased outputs without retraining the foundation model. Which TWO actions should they take?
- A
Ground the model with retrieval-augmented generation (RAG) using approved company policy documents, and instruct the model to answer only from retrieved context or say it does not know.
- B
Increase the model temperature so it can generate more diverse responses and avoid repeating potentially biased wording.
- C
Add prompt instructions that require neutral, policy-based language and prohibit unsupported assumptions about customers or products.
- D
Remove all product-line references from the source documents so the model has less information to misuse.
- E
Rely on the model provider's default safety settings alone, because they are sufficient to prevent enterprise-specific hallucinations and bias.
Show answer and explanation
Correct answers: A, C
Explanation
The best answer is to combine grounding and instruction-based controls: use retrieval-augmented generation over trusted company documents and add prompts that enforce neutral, evidence-based responses. These methods address the two stated risks without retraining the model. RAG is widely recognized as a leading pattern for reducing hallucinations in enterprise generative AI because it injects current, authoritative context at inference time. Prompt guardrails that require the model to avoid unsupported claims, use neutral language, and abstain when evidence is missing are also common best practices for reducing biased or misleading outputs. By contrast, increasing temperature usually increases variability and can worsen factual drift, while relying only on default safety features is not enough for domain-specific accuracy and fairness requirements. These practices align with common Snowflake and industry guidance for building trustworthy GenAI applications: ground responses in trusted data, constrain generation behavior, and evaluate outputs for quality and safety.
- A. Correct.
Correct. Using RAG with curated enterprise documents is a primary technique for reducing hallucinations because it grounds responses in authoritative data instead of relying only on the model's pretraining. Adding explicit instructions to answer only from retrieved context and to abstain when evidence is missing further reduces fabricated answers. This is a practical control for policy and support use cases where factual accuracy matters.
- B. Incorrect.
Incorrect. Raising temperature generally increases randomness and creativity, which can make outputs less deterministic and may increase hallucination risk rather than reduce it. It also does not directly address bias. For enterprise Q&A scenarios, lower temperature and stronger grounding are typically better choices when factual consistency is required.
- C. Correct.
Correct. Prompting for neutral, policy-based wording and explicitly forbidding unsupported assumptions is a practical mitigation for biased language when retraining is not an option. This helps steer the model toward fairer, evidence-based responses. While prompt instructions alone are not sufficient for all risk reduction, they are an important control when combined with grounding and evaluation.
- D. Incorrect.
Incorrect. Removing relevant product-line information from source documents would likely reduce answer quality and completeness, not improve trustworthiness. The better practice is to keep accurate source material and constrain the model to use it appropriately. Bias mitigation should focus on prompt design, evaluation, and curated data, not arbitrary deletion of relevant facts.
- E. Incorrect.
Incorrect. Provider safety settings can help with broad harmful-content controls, but they do not guarantee protection against organization-specific hallucinations or subtle bias in a company's domain. Enterprise applications usually require additional controls such as grounding, prompt constraints, output evaluation, and governance over approved data sources.