llm（LLM 搜索）

LLM (Large Language Model) Skill

This skill guides the implementation of chat completions functionality using the z-ai-web-dev-sdk package, enabling powerful conversational AI and text generation capabilities.

Skills Path

Skill Location: {project_path}/skills/llm

this skill is located at above path in your project.

Reference Scripts: Example test scripts are available in the {Skill Location}/scripts/ directory for quick testing and reference. See {Skill Location}/scripts/chat.ts for a working example.

Overview

The LLM skill allows you to build applications that leverage large language models for natural language understanding and generation, including chatbots, AI assistants, content generation, and more.

IMPORTANT: z-ai-web-dev-sdk MUST be used in backend code only. Never use it in client-side code.

Prerequisites

The z-ai-web-dev-sdk package is already installed. Import it as shown in the examples below.

CLI Usage (For Simple Tasks)

For simple, one-off chat completions, you can use the z-ai CLI instead of writing code. This is ideal for quick tests, simple queries, or automation scripts.

Basic Chat

# Simple question
z-ai chat --prompt "What is the capital of France?"

# Save response to file
z-ai chat -p "Explain quantum computing" -o response.json

# Stream the response
z-ai chat -p "Write a short poem" --stream

With System Prompt

# Custom system prompt for specific behavior
z-ai chat \
  --prompt "Review this code: function add(a,b) { return a+b; }" \
  --system "You are an expert code reviewer" \
  -o review.json

With Thinking (Chain of Thought)

# Enable thinking for complex reasoning
z-ai chat \
  --prompt "Solve this math problem: If a train travels 120km in 2 hours, what's its speed?" \
  --thinking \
  -o solution.json

CLI Parameters

• --prompt, -p <text>: Required - User message content
• --system, -s <text>: Optional - System prompt for custom behavior
• --thinking, -t: Optional - Enable chain-of-thought reasoning (default: disabled)
• --output, -o <path>: Optional - Output file path (JSON format)
• --stream: Optional - Stream the response in real-time

When to Use CLI vs SDK

Use CLI for:

• Quick one-off questions
• Simple automation scripts
• Testing prompts
• Single-turn conversations

Use SDK for:

• Multi-turn conversations with context
• Custom conversation management
• Integration with web applications
• Complex chat workflows
• Production applications

Basic Chat Completions

Simple Question and Answer

import ZAI from 'z-ai-web-dev-sdk';

async function askQuestion(question) {
  const zai = await ZAI.create();

  const completion = await zai.chat.completions.create({
    messages: [
      {
        role: 'assistant',
        content: 'You are a helpful assistant.'
      },
      {
        role: 'user',
        content: question
      }
    ],
    thinking: { type: 'disabled' }
  });

  const response = completion.choices[0]?.message?.content;
  return response;
}

// Usage
const answer = await askQuestion('What is the capital of France?');
console.log('Answer:', answer);

Custom System Prompt

import ZAI from 'z-ai-web-dev-sdk';

async function customAssistant(systemPrompt, userMessage) {
  const zai = await ZAI.create();

  const completion = await zai.chat.completions.create({
    messages: [
      {
        role: 'assistant',
        content: systemPrompt
      },
      {
        role: 'user',
        content: userMessage
      }
    ],
    thinking: { type: 'disabled' }
  });

  return completion.choices[0]?.message?.content;
}

// Usage - Code reviewer
const codeReview = await customAssistant(
  'You are an expert code reviewer. Analyze code for bugs, performance issues, and best practices.',
  'Review this function: function add(a, b) { return a + b; }'
);

// Usage - Creative writer
const story = await customAssistant(
  'You are a creative fiction writer who writes engaging short stories.',
  'Write a short story about a robot learning to paint.'
);

console.log(codeReview);
console.log(story);

Multi-turn Conversations

Conversation History Management

import ZAI from 'z-ai-web-dev-sdk';

class ConversationManager {
  constructor(systemPrompt = 'You are a helpful assistant.') {
    this.messages = [
      {
        role: 'assistant',
        content: systemPrompt
      }
    ];
    this.zai = null;
  }

  async initialize() {
    this.zai = await ZAI.create();
  }

  async sendMessage(userMessage) {
    // Add user message to history
    this.messages.push({
      role: 'user',
      content: userMessage
    });

    // Get completion
    const completion = await this.zai.chat.completions.create({
      messages: this.messages,
      thinking: { type: 'disabled' }
    });

    const assistantResponse = completion.choices[0]?.message?.content;

    // Add assistant response to history
    this.messages.push({
      role: 'assistant',
      content: assistantResponse
    });

    return assistantResponse;
  }

  getHistory() {
    return this.messages;
  }

  clearHistory(systemPrompt = 'You are a helpful assistant.') {
    this.messages = [
      {
        role: 'assistant',
        content: systemPrompt
      }
    ];
  }

  getMessageCount() {
    // Subtract 1 for system message
    return this.messages.length - 1;
  }
}

// Usage
const conversation = new ConversationManager();
await conversation.initialize();

const response1 = await conversation.sendMessage('Hi, my name is John.');
console.log('AI:', response1);

const response2 = await conversation.sendMessage('What is my name?');
console.log('AI:', response2); // Should remember the name is John

console.log('Total messages:', conversation.getMessageCount());

Context-Aware Conversations

import ZAI from 'z-ai-web-dev-sdk';

class ContextualChat {
  constructor() {
    this.messages = [];
    this.zai = null;
  }

  async initialize() {
    this.zai = await ZAI.create();
  }

  async startConversation(role, context) {
    // Set up system prompt with context
    const systemPrompt = `You are ${role}. Context: ${context}`;

    this.messages = [
      {
        role: 'assistant',
        content: systemPrompt
      }
    ];
  }

  async chat(userMessage) {
    this.messages.push({
      role: 'user',
      content: userMessage
    });

    const completion = await this.zai.chat.completions.create({
      messages: this.messages,
      thinking: { type: 'disabled' }
    });

    const response = completion.choices[0]?.message?.content;

    this.messages.push({
      role: 'assistant',
      content: response
    });

    return response;
  }
}

// Usage - Customer support scenario
const support = new ContextualChat();
await support.initialize();

await support.startConversation(
  'a customer support agent for TechCorp',
  'The user has ordered product #12345 which is delayed due to shipping issues.'
);

const reply1 = await support.chat('Where is my order?');
console.log('Support:', reply1);

const reply2 = await support.chat('Can I get a refund?');
console.log('Support:', reply2);

Advanced Use Cases

Content Generation

import ZAI from 'z-ai-web-dev-sdk';

class ContentGenerator {
  constructor() {
    this.zai = null;
  }

  async initialize() {
    this.zai = await ZAI.create();
  }

  async generateBlogPost(topic, tone = 'professional') {
    const completion = await this.zai.chat.completions.create({
      messages: [
        {
          role: 'assistant',
          content: `You are a professional content writer. Write in a ${tone} tone.`
        },
        {
          role: 'user',
          content: `Write a blog post about: ${topic}. Include an introduction, main points, and conclusion.`
        }
      ],
      thinking: { type: 'disabled' }
    });

    return completion.choices[0]?.message?.content;
  }

  async generateProductDescription(productName, features) {
    const completion = await this.zai.chat.completions.create({
      messages: [
        {
          role: 'assistant',
          content: 'You are an expert at writing compelling product descriptions for e-commerce.'
        },
        {
          role: 'user',
          content: `Write a product description for "${productName}". Key features: ${features.join(', ')}.`
        }
      ],
      thinking: { type: 'disabled' }
    });

    return completion.choices[0]?.message?.content;
  }

  async generateEmailResponse(originalEmail, intent) {
    const completion = await this.zai.chat.completions.create({
      messages: [
        {
          role: 'assistant',
          content: 'You are a professional email writer. Write clear, concise, and polite emails.'
        },
        {
          role: 'user',
          content: `Original email: "${originalEmail}"\n\nWrite a ${intent} response.`
        }
      ],
      thinking: { type: 'disabled' }
    });

    return completion.choices[0]?.message?.content;
  }
}

// Usage
const generator = new ContentGenerator();
await generator.initialize();

const blogPost = await generator.generateBlogPost(
  'The Future of Artificial Intelligence',
  'informative'
);
console.log('Blog Post:', blogPost);

const productDesc = await generator.generateProductDescription(
  'Smart Watch Pro',
  ['Heart rate monitoring', 'GPS tracking', 'Waterproof', '7-day battery life']
);
console.log('Product Description:', productDesc);

Data Analysis and Summarization

import ZAI from 'z-ai-web-dev-sdk';

async function analyzeData(data, analysisType) {
  const zai = await ZAI.create();

  const prompts = {
    summarize: 'You are a data analyst. Summarize the key insights from the data.',
    trend: 'You are a data analyst. Identify trends and patterns in the data.',
    recommendation: 'You are a business analyst. Provide actionable recommendations based on the data.'
  };

  const completion = await zai.chat.completions.create({
    messages: [
      {
        role: 'assistant',
        content: prompts[analysisType] || prompts.summarize
      },
      {
        role: 'user',
        content: `Analyze this data:\n\n${JSON.stringify(data, null, 2)}`
      }
    ],
    thinking: { type: 'disabled' }
  });

  return completion.choices[0]?.message?.content;
}

// Usage
const salesData = {
  Q1: { revenue: 100000, customers: 250 },
  Q2: { revenue: 120000, customers: 280 },
  Q3: { revenue: 150000, customers: 320 },
  Q4: { revenue: 180000, customers: 380 }
};

const summary = await analyzeData(salesData, 'summarize');
const trends = await analyzeData(salesData, 'trend');
const recommendations = await analyzeData(salesData, 'recommendation');

console.log('Summary:', summary);
console.log('Trends:', trends);
console.log('Recommendations:', recommendations);

Code Generation and Debugging

import ZAI from 'z-ai-web-dev-sdk';

class CodeAssistant {
  constructor() {
    this.zai = null;
  }

  async initialize() {
    this.zai = await ZAI.create();
  }

  async generateCode(description, language) {
    const completion = await this.zai.chat.completions.create({
      messages: [
        {
          role: 'assistant',
          content: `You are an expert ${language} programmer. Write clean, efficient, and well-commented code.`
        },
        {
          role: 'user',
          content: `Write ${language} code to: ${description}`
        }
      ],
      thinking: { type: 'disabled' }
    });

    return completion.choices[0]?.message?.content;
  }

  async debugCode(code, issue) {
    const completion = await this.zai.chat.completions.create({
      messages: [
        {
          role: 'assistant',
          content: 'You are an expert debugger. Identify bugs and suggest fixes.'
        },
        {
          role: 'user',
          content: `Code:\n${code}\n\nIssue: ${issue}\n\nFind the bug and suggest a fix.`
        }
      ],
      thinking: { type: 'disabled' }
    });

    return completion.choices[0]?.message?.content;
  }

  async explainCode(code) {
    const completion = await this.zai.chat.completions.create({
      messages: [
        {
          role: 'assistant',
          content: 'You are a programming teacher. Explain code clearly and simply.'
        },
        {
          role: 'user',
          content: `Explain what this code does:\n\n${code}`
        }
      ],
      thinking: { type: 'disabled' }
    });

    return completion.choices[0]?.message?.content;
  }
}

// Usage
const codeAssist = new CodeAssistant();
await codeAssist.initialize();

const newCode = await codeAssist.generateCode(
  'Create a function that sorts an array of objects by a specific property',
  'JavaScript'
);
console.log('Generated Code:', newCode);

const bugFix = await codeAssist.debugCode(
  'function add(a, b) { return a - b; }',
  'This function should add numbers but returns wrong results'
);
console.log('Debug Suggestion:', bugFix);

Best Practices

1. Prompt Engineering

// Bad: Vague prompt
const bad = await askQuestion('Tell me about AI');

// Good: Specific and structured prompt
async function askWithContext(topic, format, audience) {
  const zai = await ZAI.create();

  const completion = await zai.chat.completions.create({
    messages: [
      {
        role: 'assistant',
        content: `You are an expert educator. Explain topics clearly for ${audience}.`
      },
      {
        role: 'user',
        content: `Explain ${topic} in ${format} format. Include practical examples.`
      }
    ],
    thinking: { type: 'disabled' }
  });

  return completion.choices[0]?.message?.content;
}

const good = await askWithContext('artificial intelligence', 'bullet points', 'beginners');

2. Error Handling

import ZAI from 'z-ai-web-dev-sdk';

async function safeCompletion(messages, retries = 3) {
  let lastError;

  for (let attempt = 1; attempt <= retries; attempt++) {
    try {
      const zai = await ZAI.create();

      const completion = await zai.chat.completions.create({
        messages: messages,
        thinking: { type: 'disabled' }
      });

      const response = completion.choices[0]?.message?.content;

      if (!response || response.trim().length === 0) {
        throw new Error('Empty response from AI');
      }

      return {
        success: true,
        content: response,
        attempts: attempt
      };
    } catch (error) {
      lastError = error;
      console.error(`Attempt ${attempt} failed:`, error.message);

      if (attempt < retries) {
        // Wait before retry (exponential backoff)
        await new Promise(resolve => setTimeout(resolve, 1000 * attempt));
      }
    }
  }

  return {
    success: false,
    error: lastError.message,
    attempts: retries
  };
}

3. Context Management

class ManagedConversation {
  constructor(maxMessages = 20) {
    this.maxMessages = maxMessages;
    this.systemPrompt = '';
    this.messages = [];
    this.zai = null;
  }

  async initialize(systemPrompt) {
    this.zai = await ZAI.create();
    this.systemPrompt = systemPrompt;
    this.messages = [
      {
        role: 'assistant',
        content: systemPrompt
      }
    ];
  }

  async chat(userMessage) {
    // Add user message
    this.messages.push({
      role: 'user',
      content: userMessage
    });

    // Trim old messages if exceeding limit (keep system prompt)
    if (this.messages.length > this.maxMessages) {
      this.messages = [
        this.messages[0], // Keep system prompt
        ...this.messages.slice(-(this.maxMessages - 1))
      ];
    }

    const completion = await this.zai.chat.completions.create({
      messages: this.messages,
      thinking: { type: 'disabled' }
    });

    const response = completion.choices[0]?.message?.content;

    this.messages.push({
      role: 'assistant',
      content: response
    });

    return response;
  }

  getTokenEstimate() {
    // Rough estimate: ~4 characters per token
    const totalChars = this.messages
      .map(m => m.content.length)
      .reduce((a, b) => a + b, 0);
    return Math.ceil(totalChars / 4);
  }
}

4. Response Processing

async function getStructuredResponse(query, format = 'json') {
  const zai = await ZAI.create();

  const formatInstructions = {
    json: 'Respond with valid JSON only. No additional text.',
    list: 'Respond with a numbered list.',
    markdown: 'Respond in Markdown format.'
  };

  const completion = await zai.chat.completions.create({
    messages: [
      {
        role: 'assistant',
        content: `You are a helpful assistant. ${formatInstructions[format]}`
      },
      {
        role: 'user',
        content: query
      }
    ],
    thinking: { type: 'disabled' }
  });

  const response = completion.choices[0]?.message?.content;

  // Parse JSON if requested
  if (format === 'json') {
    try {
      return JSON.parse(response);
    } catch (e) {
      console.error('Failed to parse JSON response');
      return { raw: response };
    }
  }

  return response;
}

// Usage
const jsonData = await getStructuredResponse(
  'List three programming languages with their primary use cases',
  'json'
);
console.log(jsonData);

Common Use Cases

1. Chatbots & Virtual Assistants: Build conversational interfaces for customer support
2. Content Generation: Create articles, product descriptions, marketing copy
3. Code Assistance: Generate, explain, and debug code
4. Data Analysis: Analyze and summarize complex data sets
5. Language Translation: Translate text between languages
6. Educational Tools: Create tutoring and learning applications
7. Email Automation: Generate professional email responses
8. Creative Writing: Story generation, poetry, and creative content

Integration Examples

Express.js Chatbot API

import express from 'express';
import ZAI from 'z-ai-web-dev-sdk';

const app = express();
app.use(express.json());

// Store conversations in memory (use database in production)
const conversations = new Map();

let zaiInstance;

async function initZAI() {
  zaiInstance = await ZAI.create();
}

app.post('/api/chat', async (req, res) => {
  try {
    const { sessionId, message, systemPrompt } = req.body;

    if (!message) {
      return res.status(400).json({ error: 'Message is required' });
    }

    // Get or create conversation history
    let history = conversations.get(sessionId) || [
      {
        role: 'assistant',
        content: systemPrompt || 'You are a helpful assistant.'
      }
    ];

    // Add user message
    history.push({
      role: 'user',
      content: message
    });

    // Get completion
    const completion = await zaiInstance.chat.completions.create({
      messages: history,
      thinking: { type: 'disabled' }
    });

    const aiResponse = completion.choices[0]?.message?.content;

    // Add AI response to history
    history.push({
      role: 'assistant',
      content: aiResponse
    });

    // Save updated history
    conversations.set(sessionId, history);

    res.json({
      success: true,
      response: aiResponse,
      messageCount: history.length - 1
    });
  } catch (error) {
    res.status(500).json({
      success: false,
      error: error.message
    });
  }
});

app.delete('/api/chat/:sessionId', (req, res) => {
  const { sessionId } = req.params;
  conversations.delete(sessionId);
  res.json({ success: true, message: 'Conversation cleared' });
});

initZAI().then(() => {
  app.listen(3000, () => {
    console.log('Chatbot API running on port 3000');
  });
});

Troubleshooting

Issue: "SDK must be used in backend"

• Solution: Ensure z-ai-web-dev-sdk is only imported and used in server-side code

Issue: Empty or incomplete responses

• Solution: Check that completion.choices[0]?.message?.content exists and is not empty

Issue: Conversation context getting too long

• Solution: Implement message trimming to keep only recent messages

Issue: Inconsistent responses

• Solution: Use more specific system prompts and provide clear instructions

Issue: Rate limiting errors

• Solution: Implement retry logic with exponential backoff

Performance Tips

1. Reuse SDK Instance: Create ZAI instance once and reuse across requests
2. Manage Context Length: Trim old messages to avoid token limits
3. Implement Caching: Cache responses for common queries
4. Use Specific Prompts: Clear prompts lead to faster, better responses
5. Handle Errors Gracefully: Implement retry logic and fallback responses

Security Considerations

1. Input Validation: Always validate and sanitize user input
2. Rate Limiting: Implement rate limits to prevent abuse
3. API Key Protection: Never expose SDK credentials in client-side code
4. Content Filtering: Filter sensitive or inappropriate content
5. Session Management: Implement proper session handling and cleanup

Remember

• Always use z-ai-web-dev-sdk in backend code only
• The SDK is already installed - import as shown in examples
• Use the 'assistant' role for system prompts
• Set thinking to {type: 'disabled'} for standard completions
• Implement proper error handling and retries for production
• Manage conversation history to avoid token limits
• Clear and specific prompts lead to better results
• Check scripts/chat.ts for a quick start example