AI Integration

​

Overview

​

Technology Stack

​

Supported Models

• GPT-4o, GPT-4o-mini — OpenAI
• Claude 3.5 Sonnet, Claude 3 Haiku — Anthropic
• Gemini 2.0 Flash, Gemini 1.5 Pro — Google
• Llama 3.1, Mistral Large — Open source
• DALL-E 3, Stable Diffusion — Image generation

​

Capabilities

• Chat Completions — Text generation with streaming support
• Tool Calling — Function calling for agentic workflows
• Image Generation — Create images from text prompts
• Embeddings — Generate vector embeddings for similarity search

​

Chat Completions

​

Request

POST /api/ai/chat
{
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "Hello!" }
  ],
  "stream": true
}

​

Streaming Architecture

1. Request validated and authorized
2. Forwarded to OpenRouter
3. Tokens streamed back to client
4. Usage tracked on completion

​

Tool Calling

POST /api/ai/chat
{
  "messages": [...],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather",
        "parameters": {
          "type": "object",
          "properties": {
            "location": { "type": "string" }
          }
        }
      }
    }
  ]
}

​

Image Generation

POST /api/ai/images
{
  "prompt": "A sunset over mountains",
  "model": "dall-e-3",
  "size": "1024x1024",
  "quality": "standard"
}

​

Image Parameters

​

Configuration Management

# List configurations
GET /api/admin/ai/configs

# Create configuration
POST /api/admin/ai/configs
{
  "name": "gpt4",
  "model": "openai/gpt-4o",
  "system_prompt": "You are a helpful assistant.",
  "temperature": 0.7,
  "max_tokens": 4096
}

​

Usage Tracking

​

Metrics Collected

​

Security & Rate Limiting

• API key required for all AI requests
• Per-project rate limits based on subscription plan
• Usage quotas with configurable alerts
• No data retention — requests are not stored by Yboard

​

Best Practices

1. Use streaming for chat interfaces to improve perceived performance
2. Set system prompts in configurations rather than per-request
3. Monitor usage regularly to stay within budget
4. Use appropriate models — smaller models for simple tasks, larger for complex ones
5. Implement retry logic for transient failures