Beginner LevelAI, LLMs, Embeddings & MCP
What causes LLM Hallucinations and how do you mitigate un-grounded responses?
❓ Interview Prompt: Explain why LLMs hallucinate false information, data distribution gaps, and operational techniques to reduce hallucinations.
An LLM Hallucination occurs when a model generates text that sounds authoritative and grammatically correct, but is factually inaccurate, ungrounded, or contradictory.
Root Causes:
1
Probabilistic Autoregression: Models predict the most likely next word, not factual truth.
2
Out-of-Date / Out-of-Domain Training Data: Gaps in pre-training corpora cause models to extrapolate plausible fabrications.
Effective Mitigation Strategies:
•
Retrieval-Augmented Generation (RAG): Force the model to quote directly from retrieved ground truth documents.
•
Lower Temperature: Reduce temperature to
0.0 for factual tasks.•
System Instructions: Instruct the model: "If the answer cannot be found in the provided context, state 'I do not know'".
Code Implementation Example
typescript
const antiHallucinationSystemPrompt = "You are a strict technical assistant.\n" +
"Rules:\n" +
"1. Answer questions ONLY using the provided facts in the Context section.\n" +
"2. If the context does not contain the answer, reply EXACTLY with: 'Information not found in context.'\n" +
"3. Do NOT use outside knowledge or make assumptions.";💡Anti-hallucination system prompt enforcing strict context grounding.
🎯 Interview Key Takeaway
Mitigate hallucinations by using RAG, setting temperature to 0.0, and instructing models to refuse ungrounded questions.
#Hallucination#RAG#Prompting#Beginner
Back to all AI, LLMs, Embeddings & MCP Questions→