RAG Prompt Optimization

Expert

Engineer context prompts to minimize hallucination in vector search systems.

01 The Concept

RAG prompts integrate retrieved chunks into the LLM's context window. To prevent the model from ignoring context ('lost in the middle') or hallucinating outside the retrieved facts, prompts must define strict grounding boundaries, prioritize chunk ordering, and instruct the model to provide citations for every claim.

02 Weak vs. Strong

EX 01Strict RAG Prompt Guard

You are a customer support agent. Answer the question using ONLY the facts present in the <context_blocks> below.

<context_blocks> [Doc 1]: Refund policy: customers have 14 days to request a refund. </context_blocks>

Grounding Rules:

If the answer cannot be found in <context_blocks>, reply: 'I cannot find that in the knowledge base.'
Do NOT use any pre-trained external facts to answer the question.
Quote the Doc number (e.g. [Doc 1]) for every fact you mention.

Question: Can I get a refund after 20 days?

→ Why it works

Enforces grounding. Commands refusal on missing facts, prohibits external knowledge, and requires document citations, reducing hallucination.

03 Key Points

01Grounding boundaries: Command the model to refuse queries if the answer is missing in retrieved blocks.

02Citation enforcement: Require the model to link every claim back to a specific document or quote.

03Lost in the Middle: Put your highest-priority retrieved documents at the beginning or very end of the context block.

04Irrelevant Chunk Filtering: Instruct the model to discount noisy, irrelevant chunks during evaluation.

05Context Delimitation: Package retrieved chunks inside distinct XML tags or JSON arrays.

04 Model-Specific Notes

Claude handles multi-document context very well. Order documents using XML tags like for stable grounding and citations.

05 For Your Role

It's like taking an open-book test where you are strictly graded on quoting page numbers and forbidden from guessing.

Next: Multi-Lingual Calibration →