Senior Prompt Engineer

Root Cause: The model is inferring answers not in context — classic RAG grounding failure.

Fix 1 — Strict grounding system prompt:

SYSTEM_PROMPT = """You are a support assistant. Answer questions using ONLY the provided context.

Rules:
- If the context does not contain enough information to answer, say exactly: "I don't have enough information in our documentation to answer that. Please contact support at support@company.com."
- Do NOT use your general knowledge to fill gaps.
- Quote the relevant documentation section when answering.
- Never say 'based on my knowledge' or similar phrases."""

Fix 2 — Add a relevance gate before generation:

from langchain_core.output_parsers import JsonOutputParser

RELEVANCE_PROMPT = """Does this context contain enough information to answer the question?
Question: {question}
Context: {context}
Respond with JSON: {{"relevant": true/false, "reason": "one sentence"}}"""

def answer_with_gate(question: str, context: str) -> str:
    gate = llm.invoke(RELEVANCE_PROMPT.format(question=question, context=context))
    result = JsonOutputParser().parse(gate.content)
    
    if not result['relevant']:
        return "I don't have documentation on that topic. Please contact support."
    
    return llm.invoke(f"{SYSTEM_PROMPT}\n\nContext: {context}\n\nQuestion: {question}").content

Fix 3 — Improve retrieval (garbage in, hallucinations out):

• Use top_k=6 with MMR reranking instead of top_k=3 cosine similarity — reduces the chance the relevant chunk is missed
• Add chunk metadata (doc title, section) so the model can reference its source

Measurement: Sample 50 questions with known answers, score before/after. Grounding prompt alone typically cuts hallucination rate by 60-70%.