EXAMPLES: 1."How do I reset my password?"→category:support,urgency:low 2."Where is the password reset option?"→category:support,urgency:low 3."I can't find password settings."→category:support,urgency:low
全是同一种问题。模型学不到边界处理。
良好的 few-shot（多样且含 edge cases）：
Example 1 (Standard): Input: "How do I reset my password?" Output: {"category": "support", "urgency": "low", "confidence": 0.95}
Example 2 (Urgent): Input: "URGENT: System down, can't access customer data!" Output: {"category": "support", "urgency": "high", "confidence": 0.98}
Example 3 (Mixed Intent): Input: "I want to upgrade my plan but also report a billing error." Output: {"category": "billing", "urgency": "medium", "confidence": 0.78, "note": "Multiple intents detected"}
Example 4 (Edge Case - Unclear): Input: "help" Output: {"category": "general", "urgency": "low", "confidence": 0.45, "action": "request_clarification"}
为何有效：
覆盖不同场景（标准、紧急、混合、不清楚）
示范如何处理边界（低置信度 → 追问澄清）
展示一致的输出格式
让模型学习决策模式，而非仅做类别匹配
技巧 4：Constraints & Grounding（对抗幻觉）
AI agents 的大问题之一：hallucination（幻觉）。找不到答案时它们会编造。
解决方案：显式约束，将 agent “落地”。
糟糕做法（无约束）：
Answer customer support questions based on our documentation.
后果：找不到信息时 agent 会胡编。
良好做法（显式约束）：
Answer customer support questions using ONLY information from the documentation you can access via search_docs tool. CONSTRAINTS: - If information isnotin docs: "I don't have that information in our current documentation. I'll create a ticket for our team to help you." - Never make assumptions about features or functionality - Never provide workarounds that aren't documented - If multiple solutions exist: Present all documented options ESCALATION CRITERIA: - Customer mentions "urgent", "broken", "down" → create ticket immediately - Question requires account-specific data → create ticket with details - Documentation is incomplete/contradictory → create ticket noting the issue
You are a helpful AI assistant designed to help customers with their questions and concerns. You should always be polite, professional, and courteous in your responses. Make sure to read the customer's question carefully and provide a thorough and complete answer that addresses all of their concerns. If you're not sure about something, it's better to say you don't know than to provide incorrect information...
350 个 token 的空话，几乎没有可执行指导。
良好的 context（密度高、具体）：
You are Support Agent. RESPONSE REQUIREMENTS: - Max 150 words - Plain language (non-technical) - Structure: Problem acknowledgment → Solution → Next steps TOOLS: - search_docs(query) → search product documentation - create_ticket(title, priority, details) → escalate to human team WORKFLOW: 1.Search docs for relevant information 2. If found: Provide answer with doc reference 3. If not found OR customer mentions "urgent"/"broken": Create ticket
TASK: Analyze customer request and determine best resolution path. REASONING PROCESS (think step-by-step): 1. IDENTIFY: What is the core issue? (Quote specific parts of message) 2. CLASSIFY: Which category? (sales/support/billing/general) 3. ASSESS URGENCY: Time-sensitive keywords? Tone indicators? 4. CHECK PREREQUISITES: Can we resolve with available tools? 5. DECIDE: Route to appropriate handler with reasoning Think through eachstep explicitly before providing your final answer.
性能影响：
准确度提升：复杂推理任务提升 2–5%
延迟增加：20–40%（模型输出更多 tokens）
成本增加：与输出 token 增长成正比
结论：只有当准确度提升能抵消成本和延迟的权衡时，才使用 CoT。
模式 2：RAG（Retrieval-Augmented Generation）——用于外部知识
当你的 agent 需要：
动态/频繁更新的内容（产品目录、文档）
体量巨大、放不进 context 的知识库
客户特定数据（订单记录、账户详情）
训练语料之外的专有信息
RAG 就是必需的。
n8n 中的 RAG 基本流程：
Webhook/Trigger ↓ Extract Query (user's question) ↓ Vector Search (retrieve relevant chunks from knowledge base) ↓ AI Agent (answer using retrieved context) ↓ Response
关键 RAG 要点（基于 kapa.ai 的分析）：
Chunk size：每块 500–800 tokens（大多任务的最佳区间）
Overlap：块间重叠 50–100 tokens（避免边界信息丢失）
Number of chunks：返回 3–5 个最相关块
Reranking：向量召回后做语义重排以提升相关性
Metadata：包含来源、时间戳、置信度
RAG Prompt 示例：
Answer the customer's question using ONLY the information provided below. CONTEXT FROM DOCUMENTATION: {{$json.retrieved_chunks}} CUSTOMER QUESTION: {{$json.user_message}} INSTRUCTIONS: - Base answer strictly on provided context - If context doesn't contain the answer: "I don't have that information in our current documentation." - Include source reference: "According to [doc_title]..." - If multiple relevant sections: Synthesize information fromall CONFIDENCE ASSESSMENT: - High confidence: Answer directly stated in context - Medium confidence: Answer can be inferred from context - Low confidence: Context is incomplete → escalate
模式 3：Document Repacking——顺序比你想的更重要
Wang 等（2024）研究发现：context 的“顺序”影响显著。
发现要点：
Primacy bias：模型更注意开头的信息
Recency bias：也更注意结尾的信息
Middle neglect：中间的信息更容易被忽略
性能影响：通过最优排序可提升 5–10% 准确度
最优排序策略：
最相关/最重要的信息放最前
次要的支持信息放中间
约束与提醒放最后（利用近因效应）
示例（RAG context）：
MOST RELEVANT DOCUMENTATION: [Chunk with highest relevance score] ADDITIONAL CONTEXT: [Supporting chunks] CONSTRAINTS (IMPORTANT): - Answer only from provided context - If uncertain: Escalate to human team
I'm building an AI agent for customer support email classification. Help me create an optimal system message prompt. REQUIREMENTS: - Classify emails into: sales, support, billing, general - Assess urgency: low, medium, high - Output format: JSON with category, urgency, confidence - Must handle edge cases: unclear intent, multiple topics, spam TOOLS AVAILABLE: - search_docs(query): Search documentation - create_ticket(title, priority, description): Escalate to humans EXAMPLES OF DESIRED BEHAVIOR: [Include 3-5 diverse examples with input and expected output] CONSTRAINTS: - Never make up information - When uncertain (confidence < 0.7): Escalate - Response under 150 words for direct answers - Include reasoning in output Generate an optimized system message that will consistently produce these results.
模型会生成一个：
结构与措辞最优
融合有效技巧
在具体与灵活间取得平衡
针对你用例的 prompt
为何有效：
模型了解自己的“偏好”
它会采用最优结构和表述
你能省下大量试错时间
模型相关注意事项：哪些真的重要
大多数“模型特定技巧”并不靠谱。但有些差异确实重要：
Claude（Anthropic）：
优势：复杂推理、长 context（200K tokens）
劣势：有时过度谨慎，会拒绝无害请求
最佳实践：明确写清 constraints，再让其自由推理
Prompt caching：对 >1024 tokens 的 System Messages 启用
GPT-4o（OpenAI）：
优势：Structured Outputs（100% schema 合规）、速度快
劣势：context 较短（128K），较少“深思熟虑”
最佳实践：数据抽取使用 Structured Outputs，配合精确指令
Prompt caching：对 System Messages 自动启用
GPT-4o-mini：
优势：便宜（$0.15/M vs $3/M），适合简单任务
劣势：复杂指令鲁棒性较弱
最佳实践：使用具体、结构化的 prompts，配 few-shot 示例
Gemini（Google）：
优势：多模态（图像、视频）、超长 context（2M tokens）
劣势：tool-use 支持较弱，有时不稳定
最佳实践：用于多模态场景，避免复杂工具编排
选型经验法则：
复杂推理 + 长文档 → Claude Sonnet
数据抽取 + Structured Outputs → GPT-4o
简单分类 + 预算敏感 → GPT-4o-mini
多模态（图/视频）→ Gemini
生产级模式：测试、错误处理、优化
好 prompt 远远不够——你需要生产级工作流。
测试策略
用真实的 edge cases 测，别只测“快乐路径”：
Test cases for email triager: ✓ Standard support request ✓ Angry customer(caps, exclamation marks) ✓ Sales inquiry with technical questions(mixed intent) ✓ Very shortmessage("help") ✓ Wrong language(if only English supported) ✓ Spam/irrelevant content
错误处理
AI agents 可能失败——要有兜底：
n8n workflow: AI Agent Node → IFErrorOR confidence < 0.7: → Fallback: Route to Human → ELSE: → Continuewith automated workflow
带 confidence 的 System Message 约定：
If you're uncertain (confidence < 70%): Set"needs_human_review": truein output
粤ICP备14082021号
免费POC，
零成本试错
