✅ 最佳实践： 使用明确的分隔符或 XML 标签（Claude 和 GPT-4 对 XML 标签支持极佳）。
Here are the retrieved documents relevant to the user's question:<documents> <document index="1"> <source>employee_handbook.pdf</source> <content>报销必须在费用发生后的 30 天内提交...</content> </document> <document index="2"> <source>travel_policy.docx</source> <content>所有超过 5000 元的差旅费需要 VP 审批...</content> </document></documents>
这种结构化的注入方式有三个好处：
边界清晰：模型能明确区分不同文档。
便于引用：每个文档都有 index，方便要求模型引用。
元数据感知：可以包含 source、date 等元数据，帮助模型判断时效性。
2. 引用与溯源
RAG 的核心价值在于“有据可查”。我们需要强制模型在回答中标注来源。
提示词策略：
"Answer the user's question using only the information in the <documents>. You must cite the document index for every statement you make, using the format [index]. For example: 'Expenses must be submitted within 30 days [1].'"
"If the provided documents do not contain the necessary information to answer the question, you must say 'The provided documents do not contain this information.' Do NOT try to answer from your own knowledge. Do NOT make up an answer."
思维链强化： 要求模型在回答前先检查文档覆盖率：
"Thought: First, check if the documents contain the answer to the user's specific question. If yes, extract the info. If no, output the refusal message."
4. 答案优化与合成
综合冲突信息
当不同文档在同一事实上有出入时（如旧文档说 A，新文档说 B）：
"If there are conflicting statements in the documents, prioritize the document with the more recent date metadata. If dates are not available, mention the conflict in your answer."
风格统一
RAG 的检索结果往往风格迥异（有的来自法条，有的来自聊天记录）。提示词需要统一下一步生成的语调：
"Synthesize the information into a coherent, professional response suitable for a customer service email. Avoid copying the text verbatim; rephrase it to be more natural."
9.3.1 完整模板示例
完整 API 调用示例
以下示例展示如何将上述 RAG 提示词模板与 API 调用结合，实现一个端到端的 RAG 问答流程：
[!TIP] 在生产环境中，建议将 System Prompt（角色定义和规则）放入 system 参数，将检索文档和用户问题放入 messages，以便利用提示词缓存降低重复调用的成本。
# Role
You are an expert support assistant for [Company Name].
# Task
Answer the user's question based strictly on the provided context.
# Context
<documents>
{context_str}
</documents>
# Rules
1. **Evidence-Based**: Every sentence in your answer must be supported by the context.
2. **Citation**: Cite sources using [index] format at the end of sentences.
3. **No Hallucination**: If the answer is not in the context, say "I cannot find the answer in the knowledge base."
4. **Tone**: Helpful, concise, and professional.
# User Question
{query_str}
# Answer
“””
rag_api_example.py
使用 Anthropic Claude API 的 RAG 问答完整示例。
依赖: pip install anthropic
“””
import anthropic
# ── 1. 模拟检索结果（实际场景中来自向量数据库）──
retrieved_docs = [
{“source”: “employee_handbook.pdf”, “content”: “报销必须在费用发生后的 30 天内提交。超过 30 天的报销申请需要部门经理特批。”},
{“source”: “travel_policy.docx”, “content”: “国内差旅住宿标准：一线城市不超过 800 元/晚，其他城市不超过 500 元/晚。”},
]
# ── 2. 构造 RAG 提示词 ──
def build_rag_prompt(docs: list, query: str) -> str:
doc_xml = “\n”.join(
f' <document index=”{i+1}”>\n'
f' <source>{d[“source”]}</source>\n'
f' <content>{d[“content”]}</content>\n'
f' </document>'
for i, d in enumerate(docs)
)
return f”””Here are the retrieved documents relevant to the user's question:
<documents>
{doc_xml}
</documents>
# Rules
1. **Evidence-Based**: Every sentence must be supported by the documents above.
2. **Citation**: Cite sources using [index] format.
3. **No Hallucination**: If the answer is not in the documents, say “根据现有知识库，我无法找到相关答案。”
4. **Tone**: Helpful, concise, professional.
# User Question
{query}
# Answer”””
# ── 3. 调用 API ──
client = anthropic.Anthropic() # 使用 ANTHROPIC_API_KEY 环境变量
query = “报销的截止时间是多久？”
prompt = build_rag_prompt(retrieved_docs, query)
response = client.messages.create(
model=”claude-sonnet-4-6”,
max_tokens=1024,
messages=[{“role”: “user”, “content”: prompt}]
)
print(response.content[0].text)
# 预期输出：报销必须在费用发生后的 30 天内提交 [1]。如果超过 30 天，需要获得部门经理的特批 [1]。
