AI Integration Patterns

When to use: Any LLM integration, structured data extraction, prompt engineering, or AI-powered content analysis.

Overview

Battle-tested patterns for integrating AI models with reliable structured output parsing, multi-provider support, and comprehensive error handling.

Supported Providers

•✅ OpenAI (GPT-4, GPT-3.5)
•✅ Qwen (通义千问)
•✅ Anthropic Claude
•✅ Custom API endpoints

Key Capabilities

1. Structured Output Parsing

Problem: AI models return text, but you need JSON.

Solution: Multi-strategy parsing with automatic fallback:

javascript

function parseStructuredOutput(text) {
  // Strategy 1: Direct JSON parse
  try {
    var parsed = JSON.parse(text);
    if (parsed && typeof parsed === 'object') {
      return parsed;
    }
  } catch (e) {}

  // Strategy 2: Extract from markdown code block
  var codeBlockMatch = text.match(/```(?:json)?\s*\n([\s\S]*?)\n```/);
  if (codeBlockMatch) {
    try {
      return JSON.parse(codeBlockMatch[1]);
    } catch (e) {}
  }

  // Strategy 3: Find JSON-like content
  var jsonMatch = text.match(/\{[\s\S]*\}/);
  if (jsonMatch) {
    try {
      return JSON.parse(jsonMatch[0]);
    } catch (e) {}
  }

  throw new Error('Failed to parse structured output');
}

2. Prompt Template Building

Effective prompt structure:

javascript

function buildStructuredPrompt(content, schema) {
  var prompt = 'Analyze the following content and extract information.\n\n';
  prompt += 'Content:\n' + content + '\n\n';
  prompt += 'Return ONLY a JSON object with this exact structure:\n';
  prompt += JSON.stringify(schema, null, 2) + '\n\n';
  prompt += 'Rules:\n';
  prompt += '1. Return valid JSON only, no markdown\n';
  prompt += '2. Follow the schema exactly\n';
  prompt += '3. Use null for missing data\n';
  return prompt;
}

3. Multi-Provider Request Handler

ES5 compatible HTTP request:

javascript

var https = require('https');

function callAI(provider, prompt, apiKey) {
  var options = {
    hostname: provider === 'openai' ? 'api.openai.com' : 'dashscope.aliyuncs.com',
    path: provider === 'openai' ? '/v1/chat/completions' : '/api/v1/services/aigc/text-generation/generation',
    method: 'POST',
    headers: {
      'Content-Type': 'application/json',
      'Authorization': 'Bearer ' + apiKey
    }
  };

  var body = provider === 'openai'
    ? JSON.stringify({
        model: 'gpt-4',
        messages: [{role: 'user', content: prompt}],
        temperature: 0.1
      })
    : JSON.stringify({
        model: 'qwen-max',
        input: {prompt: prompt},
        parameters: {temperature: 0.1}
      });

  return new Promise(function(resolve, reject) {
    var req = https.request(options, function(res) {
      var data = '';
      res.on('data', function(chunk) { data += chunk; });
      res.on('end', function() {
        if (res.statusCode !== 200) {
          reject(new Error('API error: ' + res.statusCode));
        } else {
          resolve(JSON.parse(data));
        }
      });
    });
    req.on('error', reject);
    req.write(body);
    req.end();
  });
}

4. Schema Validation

javascript

function validateSchema(data, schema) {
  var errors = [];

  for (var key in schema) {
    if (!data.hasOwnProperty(key)) {
      errors.push('Missing field: ' + key);
    } else if (typeof data[key] !== typeof schema[key]) {
      errors.push('Type mismatch for ' + key);
    }
  }

  if (errors.length > 0) {
    throw new Error('Schema validation failed: ' + errors.join(', '));
  }

  return true;
}

Complete Workflow Pattern

code

Input Text
    ↓
[Build Prompt]
    - Add schema requirements
    - Add formatting rules
    - Add examples (few-shot)
    ↓
[Call AI API]
    - Select provider (OpenAI/Qwen/Claude)
    - Handle rate limits
    - Retry on errors
    ↓
[Parse Response]
    - Try direct JSON parse
    - Extract from code blocks
    - Find JSON patterns
    ↓
[Validate Schema]
    - Check required fields
    - Verify data types
    - Handle missing data
    ↓
Structured Output

n8n Implementation

Basic Structure

javascript

// Code node 1: Build Prompt
var schema = {
  title: '',
  summary: '',
  tags: [],
  category: ''
};

var prompt = buildStructuredPrompt($input.item.json.content, schema);

return {
  json: Object.assign({}, $input.item.json, {
    prompt: prompt,
    schema: schema
  })
};

javascript

// Code node 2: Call AI
var apiKey = process.env.OPENAI_API_KEY;
var response = await callAI('openai', $input.item.json.prompt, apiKey);

return {
  json: Object.assign({}, $input.item.json, {
    aiResponse: response
  })
};

javascript

// Code node 3: Parse & Validate
var text = $input.item.json.aiResponse.choices[0].message.content;
var parsed = parseStructuredOutput(text);
validateSchema(parsed, $input.item.json.schema);

return {
  json: Object.assign({}, $input.item.json, parsed)
};

Best Practices

•Low Temperature: Use 0.1-0.3 for structured output
•Explicit Instructions: "Return ONLY JSON, no markdown"
•Schema in Prompt: Show exact structure expected
•Multi-Strategy Parsing: Try multiple extraction methods
•Validate Always: Check schema before using data
•Error Handling: Retry on parse failures with adjusted prompt
•Few-Shot Examples: Include 1-2 examples for complex schemas

Common Patterns

Pattern 1: Content Analysis

code

Text → AI Analysis → Structured Data (title, summary, tags)

Pattern 2: Batch Processing

code

Multiple Items → Aggregate → Single AI Call → Split Results

Pattern 3: Multi-Step Analysis

code

Extract → Classify → Enrich → Validate → Store

Troubleshooting

AI returns markdown instead of JSON

javascript

// Add to prompt:
"CRITICAL: Return raw JSON only. No markdown formatting. No code blocks."

Inconsistent field names

javascript

// Add normalization:
var normalized = {
  title: data.title || data.Title || data.heading || '',
  // ... other fields
};

Rate limiting

javascript

// Add delay between requests
await new Promise(function(resolve) {
  setTimeout(resolve, 1000);  // 1 second
});

Integration with Other Skills

•video-processing: Analyze video transcripts
•notion-operations: Save structured analysis to Notion
•error-handling: Retry AI calls on failures
•debugging: Validate AI output quality

Full Code and Documentation

Complete implementations: /mnt/d/work/n8n_agent/n8n-skills/ai-integration/

Files:

•structured-output-handler.js - Complete parsing solution
•README.md - Detailed patterns and examples
•Prompt templates and schemas