Model Management - ElizaOS Documentation

Model System

The model management system provides a unified interface for AI model access with automatic provider selection, priority-based routing, and type-safe parameters.

Model Types

Core Model Types

enum ModelType {
  // Text generation models
  TEXT_SMALL = 'text:small',      // Fast, simple responses
  TEXT_MEDIUM = 'text:medium',    // Balanced performance
  TEXT_LARGE = 'text:large',      // Complex reasoning
  
  // Embedding models
  TEXT_EMBEDDING = 'text:embedding',
  
  // Image models
  IMAGE_GENERATION = 'image:generation',
  IMAGE_ANALYSIS = 'image:analysis',
  
  // Audio models
  SPEECH_TO_TEXT = 'speech:to:text',
  TEXT_TO_SPEECH = 'text:to:speech',
  
  // Specialized models
  CODE_GENERATION = 'code:generation',
  CLASSIFICATION = 'classification'
}

Model Parameters

Type-safe parameters for each model type:

// Text generation parameters
interface TextGenerationParams {
  prompt: string;
  messages?: Message[];
  temperature?: number;      // 0.0 - 2.0
  maxTokens?: number;
  topP?: number;
  frequencyPenalty?: number;
  presencePenalty?: number;
  stopSequences?: string[];
  systemPrompt?: string;
}

// Embedding parameters
interface EmbeddingParams {
  input: string | string[];
  model?: string;
  dimensions?: number;
}

// Image generation parameters
interface ImageGenerationParams {
  prompt: string;
  negativePrompt?: string;
  width?: number;
  height?: number;
  steps?: number;
  seed?: number;
  style?: string;
}

// Speech-to-text parameters
interface SpeechToTextParams {
  audio: Buffer | string;  // Audio data or URL
  language?: string;
  format?: 'json' | 'text' | 'srt';
  temperature?: number;
}

Model Registration

Registering Model Handlers

// Register a model handler
runtime.registerModel(
  ModelType.TEXT_LARGE,
  async (runtime, params) => {
    // Model implementation
    const response = await callAPI(params);
    return response.text;
  },
  'openai',  // provider name
  100        // priority (higher = preferred)
);

// Register multiple models from a plugin
const modelPlugin: Plugin = {
  name: 'openai-models',
  
  models: [
    {
      type: ModelType.TEXT_LARGE,
      handler: handleTextGeneration,
      provider: 'openai',
      priority: 100
    },
    {
      type: ModelType.TEXT_EMBEDDING,
      handler: handleEmbedding,
      provider: 'openai',
      priority: 100
    }
  ]
};

Model Handler Interface

type ModelHandler<T = any, R = any> = (
  runtime: IAgentRuntime,
  params: T
) => Promise<R>;

interface ModelRegistration {
  type: ModelTypeName;
  handler: ModelHandler;
  provider: string;
  priority: number;
}

Using Models

Type-Safe Model Usage

// Text generation
const response = await runtime.useModel(
  ModelType.TEXT_LARGE,
  {
    prompt: "Explain quantum computing",
    temperature: 0.7,
    maxTokens: 500
  }
);

// Get embeddings
const embedding = await runtime.useModel(
  ModelType.TEXT_EMBEDDING,
  { input: "Text to embed" }
);

// Generate image
const image = await runtime.useModel(
  ModelType.IMAGE_GENERATION,
  {
    prompt: "A sunset over mountains",
    width: 1024,
    height: 1024,
    steps: 50
  }
);

// Speech to text
const transcript = await runtime.useModel(
  ModelType.SPEECH_TO_TEXT,
  {
    audio: audioBuffer,
    language: 'en',
    format: 'json'
  }
);

Specifying Provider

// Use specific provider
const response = await runtime.useModel(
  ModelType.TEXT_LARGE,
  { prompt: "Hello" },
  'anthropic'  // Force specific provider
);

// Get available providers
const providers = runtime.getModelProviders(ModelType.TEXT_LARGE);
console.log('Available providers:', providers);
// ['openai', 'anthropic', 'ollama']

Provider Priority

Priority System

The runtime selects providers based on priority:

// Higher priority providers are preferred
runtime.registerModel(ModelType.TEXT_LARGE, handlerA, 'provider-a', 100);
runtime.registerModel(ModelType.TEXT_LARGE, handlerB, 'provider-b', 90);
runtime.registerModel(ModelType.TEXT_LARGE, handlerC, 'provider-c', 80);

// Will use provider-a (priority 100)
await runtime.useModel(ModelType.TEXT_LARGE, params);

Fallback Mechanism

// Automatic fallback on failure
class ModelRouter {
  async useModel(type: ModelType, params: any, preferredProvider?: string) {
    const providers = this.getProvidersByPriority(type, preferredProvider);
    
    for (const provider of providers) {
      try {
        return await provider.handler(this.runtime, params);
      } catch (error) {
        this.logger.warn(`Provider ${provider.name} failed:`, error);
        
        // Try next provider
        if (provider !== providers[providers.length - 1]) {
          continue;
        }
        
        // All providers failed
        throw new Error(`No providers available for ${type}`);
      }
    }
  }
}

Model Providers

OpenAI Provider

class OpenAIModelProvider {
  private client: OpenAI;
  
  constructor(runtime: IAgentRuntime) {
    const apiKey = runtime.getSetting('OPENAI_API_KEY');
    this.client = new OpenAI({ apiKey });
  }
  
  async handleTextGeneration(params: TextGenerationParams) {
    const response = await this.client.chat.completions.create({
      model: params.model || 'gpt-4',
      messages: params.messages || [
        { role: 'user', content: params.prompt }
      ],
      temperature: params.temperature,
      max_tokens: params.maxTokens,
      top_p: params.topP,
      frequency_penalty: params.frequencyPenalty,
      presence_penalty: params.presencePenalty,
      stop: params.stopSequences
    });
    
    return response.choices[0].message.content;
  }
  
  async handleEmbedding(params: EmbeddingParams) {
    const response = await this.client.embeddings.create({
      model: 'text-embedding-3-small',
      input: params.input,
      dimensions: params.dimensions
    });
    
    return Array.isArray(params.input)
      ? response.data.map(d => d.embedding)
      : response.data[0].embedding;
  }
  
  register(runtime: IAgentRuntime) {
    runtime.registerModel(
      ModelType.TEXT_LARGE,
      this.handleTextGeneration.bind(this),
      'openai',
      100
    );
    
    runtime.registerModel(
      ModelType.TEXT_EMBEDDING,
      this.handleEmbedding.bind(this),
      'openai',
      100
    );
  }
}

Anthropic Provider

class AnthropicModelProvider {
  private client: Anthropic;
  
  constructor(runtime: IAgentRuntime) {
    const apiKey = runtime.getSetting('ANTHROPIC_API_KEY');
    this.client = new Anthropic({ apiKey });
  }
  
  async handleTextGeneration(params: TextGenerationParams) {
    const response = await this.client.messages.create({
      model: params.model || 'claude-3-opus-20240229',
      messages: params.messages || [
        { role: 'user', content: params.prompt }
      ],
      max_tokens: params.maxTokens || 1000,
      temperature: params.temperature,
      system: params.systemPrompt
    });
    
    return response.content[0].text;
  }
  
  register(runtime: IAgentRuntime) {
    runtime.registerModel(
      ModelType.TEXT_LARGE,
      this.handleTextGeneration.bind(this),
      'anthropic',
      95 // Slightly lower priority than OpenAI
    );
  }
}

Local Model Provider (Ollama)

class OllamaModelProvider {
  private baseUrl: string;
  
  constructor(runtime: IAgentRuntime) {
    this.baseUrl = runtime.getSetting('OLLAMA_BASE_URL') || 'http://localhost:11434';
  }
  
  async handleTextGeneration(params: TextGenerationParams) {
    const response = await fetch(`${this.baseUrl}/api/generate`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        model: params.model || 'llama2',
        prompt: params.prompt,
        temperature: params.temperature,
        options: {
          num_predict: params.maxTokens,
          top_p: params.topP,
          stop: params.stopSequences
        }
      })
    });
    
    const data = await response.json();
    return data.response;
  }
  
  async handleEmbedding(params: EmbeddingParams) {
    const response = await fetch(`${this.baseUrl}/api/embeddings`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        model: params.model || 'all-minilm',
        prompt: params.input
      })
    });
    
    const data = await response.json();
    return data.embedding;
  }
  
  register(runtime: IAgentRuntime) {
    // Lower priority for local models
    runtime.registerModel(
      ModelType.TEXT_LARGE,
      this.handleTextGeneration.bind(this),
      'ollama',
      50
    );
    
    runtime.registerModel(
      ModelType.TEXT_EMBEDDING,
      this.handleEmbedding.bind(this),
      'ollama',
      50
    );
  }
}

Model Selection Strategy

Automatic Selection

// Runtime automatically selects best available provider
const response = await runtime.useModel(
  ModelType.TEXT_LARGE,
  { prompt: "Hello" }
);

// Selection order:
// 1. Check if preferred provider specified
// 2. Sort available providers by priority
// 3. Try each provider until success
// 4. Cache successful provider for session

Context-Based Selection

// Select model based on context
async function selectModelForTask(runtime: IAgentRuntime, task: string) {
  const complexity = analyzeComplexity(task);
  
  if (complexity < 0.3) {
    // Simple task - use small model
    return runtime.useModel(ModelType.TEXT_SMALL, {
      prompt: task,
      temperature: 0.3
    });
  } else if (complexity < 0.7) {
    // Medium complexity - use medium model
    return runtime.useModel(ModelType.TEXT_MEDIUM, {
      prompt: task,
      temperature: 0.5
    });
  } else {
    // Complex task - use large model
    return runtime.useModel(ModelType.TEXT_LARGE, {
      prompt: task,
      temperature: 0.7,
      maxTokens: 2000
    });
  }
}

Cost Optimization

// Track and optimize model usage costs
class CostOptimizedModelRouter {
  private costs = {
    'openai': { [ModelType.TEXT_LARGE]: 0.03, [ModelType.TEXT_EMBEDDING]: 0.0001 },
    'anthropic': { [ModelType.TEXT_LARGE]: 0.025 },
    'ollama': { [ModelType.TEXT_LARGE]: 0, [ModelType.TEXT_EMBEDDING]: 0 }
  };
  
  async useModel(type: ModelType, params: any, maxCost?: number) {
    const providers = this.getProvidersByCost(type, maxCost);
    
    for (const provider of providers) {
      try {
        const result = await provider.handler(this.runtime, params);
        
        // Track usage
        this.trackUsage(provider.name, type, params);
        
        return result;
      } catch (error) {
        continue;
      }
    }
  }
  
  private getProvidersByCost(type: ModelType, maxCost?: number) {
    return this.providers
      .filter(p => {
        const cost = this.costs[p.name]?.[type] || Infinity;
        return !maxCost || cost <= maxCost;
      })
      .sort((a, b) => {
        const costA = this.costs[a.name]?.[type] || Infinity;
        const costB = this.costs[b.name]?.[type] || Infinity;
        return costA - costB;
      });
  }
}

Model Caching

Response Caching

class ModelCache {
  private cache = new Map<string, { result: any; timestamp: number }>();
  private ttl = 60 * 60 * 1000; // 1 hour
  
  getCacheKey(type: ModelType, params: any): string {
    return `${type}:${JSON.stringify(params)}`;
  }
  
  get(type: ModelType, params: any): any | null {
    const key = this.getCacheKey(type, params);
    const cached = this.cache.get(key);
    
    if (!cached) return null;
    
    if (Date.now() - cached.timestamp > this.ttl) {
      this.cache.delete(key);
      return null;
    }
    
    return cached.result;
  }
  
  set(type: ModelType, params: any, result: any) {
    const key = this.getCacheKey(type, params);
    this.cache.set(key, {
      result,
      timestamp: Date.now()
    });
  }
}

// Use with runtime
const cache = new ModelCache();

async function cachedModelCall(runtime: IAgentRuntime, type: ModelType, params: any) {
  // Check cache
  const cached = cache.get(type, params);
  if (cached) return cached;
  
  // Make call
  const result = await runtime.useModel(type, params);
  
  // Cache result
  cache.set(type, params, result);
  
  return result;
}

Model Monitoring

Usage Tracking

interface ModelUsageMetrics {
  provider: string;
  modelType: ModelType;
  count: number;
  totalTokens: number;
  totalDuration: number;
  avgDuration: number;
  errors: number;
  cost: number;
}

class ModelMonitor {
  private metrics = new Map<string, ModelUsageMetrics>();
  
  async trackUsage(
    provider: string,
    type: ModelType,
    params: any,
    result: any,
    duration: number
  ) {
    const key = `${provider}:${type}`;
    
    if (!this.metrics.has(key)) {
      this.metrics.set(key, {
        provider,
        modelType: type,
        count: 0,
        totalTokens: 0,
        totalDuration: 0,
        avgDuration: 0,
        errors: 0,
        cost: 0
      });
    }
    
    const metrics = this.metrics.get(key);
    metrics.count++;
    metrics.totalDuration += duration;
    metrics.avgDuration = metrics.totalDuration / metrics.count;
    
    // Estimate tokens (simplified)
    if (type === ModelType.TEXT_LARGE) {
      const tokens = this.estimateTokens(params.prompt) + 
                     this.estimateTokens(result);
      metrics.totalTokens += tokens;
      metrics.cost += this.calculateCost(provider, type, tokens);
    }
    
    // Emit metrics event
    await this.runtime.emit(EventType.MODEL_USED, {
      runtime: this.runtime,
      modelType: type,
      provider,
      params,
      result,
      duration,
      metrics
    });
  }
}

Error Handling

Retry Logic

async function modelCallWithRetry(
  runtime: IAgentRuntime,
  type: ModelType,
  params: any,
  maxRetries = 3
) {
  let lastError: Error;
  
  for (let i = 0; i < maxRetries; i++) {
    try {
      return await runtime.useModel(type, params);
    } catch (error) {
      lastError = error;
      
      // Check if retryable
      if (isRateLimitError(error)) {
        // Wait with exponential backoff
        const delay = Math.pow(2, i) * 1000;
        await new Promise(resolve => setTimeout(resolve, delay));
        continue;
      }
      
      // Non-retryable error
      throw error;
    }
  }
  
  throw lastError;
}

Graceful Degradation

// Fallback to simpler models on failure
async function modelCallWithFallback(
  runtime: IAgentRuntime,
  params: TextGenerationParams
) {
  const modelHierarchy = [
    ModelType.TEXT_LARGE,
    ModelType.TEXT_MEDIUM,
    ModelType.TEXT_SMALL
  ];
  
  for (const modelType of modelHierarchy) {
    try {
      return await runtime.useModel(modelType, params);
    } catch (error) {
      runtime.logger.warn(`Model ${modelType} failed, trying fallback`);
      
      if (modelType === ModelType.TEXT_SMALL) {
        // Last option failed
        throw error;
      }
    }
  }
}

Best Practices

Model Selection

Right-Size Models: Use appropriate model size for task complexity
Cost Awareness: Consider cost when selecting providers
Latency Requirements: Use local models for low-latency needs
Fallback Strategy: Implement fallbacks for reliability
Caching: Cache responses for repeated queries

Performance

Batch Processing: Batch multiple requests when possible
Streaming: Use streaming for long responses
Timeout Handling: Set appropriate timeouts
Connection Pooling: Reuse HTTP connections
Rate Limiting: Respect provider rate limits

Monitoring

Track Usage: Monitor token usage and costs
Error Rates: Track provider error rates
Latency Metrics: Monitor response times
Quality Metrics: Track response quality
Cost Analysis: Analyze cost per request

What’s Next?

Services

Build services that provide models

Messaging

Stream model responses in real-time

Sessions

Use models in conversations

Providers

Supply context to models

GETTING STARTED

GUIDES

PROJECTS

AGENTS

PLUGINS

RUNTIME

​Model System

​Model Types

​Core Model Types

​Model Parameters

​Model Registration

​Registering Model Handlers

​Model Handler Interface

​Using Models

​Type-Safe Model Usage

​Specifying Provider

​Provider Priority

​Priority System

​Fallback Mechanism

​Model Providers

​OpenAI Provider

​Anthropic Provider

​Local Model Provider (Ollama)

​Model Selection Strategy

​Automatic Selection

​Context-Based Selection

​Cost Optimization

​Model Caching

​Response Caching

​Model Monitoring

​Usage Tracking

​Error Handling

​Retry Logic

​Graceful Degradation

​Best Practices

​Model Selection

​Performance

​Monitoring

​What’s Next?