SnowPro Specialty: Gen AI Question 244
Select 2Methods to reduce model hallucinations and biasA financial services company uses a large language model in a Snowflake-based application to answer employee questions about internal HR policies. During testing, the team finds two recurring problems: the model sometimes invents policy details that are not in the official documents, and its answers about parental leave vary depending on whether the question references "mother," "father," or "adoptive parent." The team wants to reduce both hallucinations and bias before production rollout. Which TWO actions would best address these issues?
- A
Ground the model with retrieval from approved HR policy documents and instruct it to answer only from retrieved context, including saying it does not know when the context is insufficient.
- B
Increase the model temperature so it can generate more diverse answers and avoid repeating potentially biased wording.
- C
Evaluate outputs with representative prompts that cover different employee groups and family situations, then refine prompts and source content based on the findings.
- D
Remove all system instructions and let the model rely on its pretrained knowledge so it can generalize across different policy questions.
- E
Use a larger context window to include unrelated corporate documents, such as marketing materials and quarterly business reviews, so the model has more information to choose from.
Show answer and explanation
Correct answers: A, C
Explanation
The best answer is to combine grounded retrieval with targeted evaluation for bias. In enterprise GenAI applications, hallucinations are commonly reduced by grounding generation in trusted documents and instructing the model to abstain when the answer is not supported by retrieved context. Bias is better addressed through representative evaluation sets and iterative refinement of prompts, guardrails, and source materials to ensure consistent treatment across protected or sensitive groups. These approaches align with widely accepted GenAI best practices: use authoritative context for factual tasks, prefer controlled generation settings for high-stakes responses, and validate outputs with realistic test cases that reflect diverse users and scenarios.
- A. Correct.
Correct. Retrieval-augmented generation grounded in approved HR documents is a primary method to reduce hallucinations because it constrains answers to authoritative sources instead of relying on the model's parametric knowledge. Adding an instruction to abstain or say it does not know when the retrieved context is insufficient further reduces fabricated answers. This is a practical best practice for enterprise GenAI systems that must provide accurate, source-based responses.
- B. Incorrect.
Incorrect. Higher temperature generally increases randomness and creativity, which tends to increase variability and can worsen hallucinations. It is not a reliable method for reducing bias either. For policy Q&A use cases, lower temperature and stronger grounding are usually better choices when consistency and factual accuracy are required.
- C. Correct.
Correct. Bias is often discovered through systematic evaluation across demographic and scenario variations, such as prompts using mother, father, adoptive parent, caregiver, or same-sex partner. Testing with representative prompts helps identify disparate behavior, while refining prompts, guardrails, and source content helps mitigate it. This is an applied governance and evaluation practice rather than a one-time model setting change.
- D. Incorrect.
Incorrect. Removing system instructions typically makes behavior less controlled, not more reliable. System instructions are useful for setting boundaries such as using only approved documents, avoiding unsupported claims, and responding consistently across policy-related scenarios. Relying on pretrained knowledge alone is especially risky for internal policy questions because those policies are organization-specific and may not exist in the model's training data.
- E. Incorrect.
Incorrect. More context is not inherently better if the added material is irrelevant. Including unrelated corporate documents can dilute signal, increase confusion, and make grounding less precise. To reduce hallucinations, retrieved context should be relevant, authoritative, and tightly scoped to the question being asked.