The Semantic Search & RAG Grounding Expert

Top-Prompts

Synthesize answers strictly from context documents.

01 The Concept

Standard Retrieval-Augmented Generation (RAG) prompts suffer from hallucinations: models merge their internal weights with context, cite missing files, or guess when documents don't contain the answer.

This template enforces factual grounding constraints, demands exact quote citations, specifies contradiction resolution rules, and implements a strict fallback rule to eliminate hallucination.

02 Weak vs. Strong

EX 01Secure Internal Doc Q&A

You are a factual Document QA Assistant. [CONTEXT DOCUMENTS] <doc id="doc_1"> TaskFlow's billing portal is accessible at billing.taskflow.app. Subscriptions renew on the 1st of the month. </doc> <doc id="doc_2"> Invoices are generated within 24 hours of payment. For refunds, contact [email protected]. </doc> [RULES] 1. Base your answer ONLY on the provided CONTEXT DOCUMENTS. Do not use outside knowledge or assumptions. 2. For every fact stated, cite the source document ID and include the exact matching quote in brackets (e.g., "...renew on the 1st [doc_1: 'renew on the 1st of the month']"). 3. Fallback: If the documents do not contain the answer, reply: "I cannot answer this question as the information is not present in the provided documents."

→ Why it works

Forbids external weights, requires bracketed quotes for all statements, and enforces a strict fallback string to block guessing.

EX 02Contradictory Policy Resolution

You are a strict RAG Synthesis Agent. [CONTEXT] <doc id="refund_2025"> Refund requests must be submitted within 14 days of purchase. </doc> <doc id="refund_2026"> We have updated our policy. All purchases are final. No refunds are issued. </doc> [TASK] Summarize our refund policy. If the sources contradict each other, do not choose one. State the contradiction, cite both source documents, and quote them.

→ Why it works

Explicitly handles contradictory documentation, forcing the model to report conflicts rather than resolving them arbitrarily.

03 Key Points

01Factual grounding: forbid the model from using outside knowledge for facts.

02Exact quote citations: require every claim to be followed by a bracketed quote from the text.

03Uncertainty fallback: command the model to state 'Not mentioned in source documents' rather than guessing.

04Discrepancy resolution: instruct the model to highlight conflicting information between sources rather than choosing one.

05JSON/Markdown structure: output answers with metadata containing citation keys.

04 Model-Specific Notes

Claude excels at RAG grounding. It adheres to 'answer only from context' rules better than most models. Wrapping documents in clear XML tags is highly recommended.

05 For Your Role

Think of it like an open-book test. The student is only allowed to write down what is written on the page. If it's not on the page, they must leave the answer blank.

Next: SQL Query Generator & Guard →