Provider System

Understanding how Runtime abstracts LLM providers.

Overview

Runtime uses a provider abstraction to work with multiple LLM services (OpenAI, Anthropic Claude, Google Gemini) through a unified interface.

Core Concept

All providers implement the same interface:

type Provider interface {
    Complete(ctx context.Context, messages []Message, config *ProviderConfig) (*ProviderResponse, error)
    CompleteStream(ctx context.Context, messages []Message, config *ProviderConfig) (StreamReader, error)
    GetProviderName() string
    Close() error
}

This allows code like:

// Same code works with any provider
result, err := provider.Complete(ctx, messages, config)

Why Provider Abstraction?

Problem: Vendor Lock-In

Without abstraction:

// Tied to OpenAI
response := openai.ChatCompletion(...)

// Want to switch to Claude? Rewrite everything
response := anthropic.Messages(...)

// Different APIs, different parameters, different response formats

Solution: Common Interface

With abstraction:

// Works with any provider
var provider types.Provider

// OpenAI
provider = openai.NewOpenAIProvider(...)

// Or Claude
provider = anthropic.NewAnthropicProvider(...)

// Or Gemini
provider = gemini.NewGeminiProvider(...)

// Same code!
response, err := provider.Complete(ctx, messages, config)

Benefits

1. Provider Independence

2. Fallback Strategies

3. Cost Optimization

4. Testing

Provider Interface

Complete Method

Synchronous completion:

Complete(ctx context.Context, messages []Message, config *ProviderConfig) (*ProviderResponse, error)

Parameters:

Returns:

CompleteStream Method

Streaming completion:

CompleteStream(ctx context.Context, messages []Message, config *ProviderConfig) (StreamReader, error)

Returns a stream reader for real-time output.

Lifecycle Methods

GetProviderName() string  // Returns "openai", "claude", "gemini"
Close() error             // Cleanup resources

Provider Configuration

Unified config works across all providers:

type ProviderConfig struct {
    MaxTokens     int      // Output limit
    Temperature   float64  // Randomness (0.0-2.0)
    TopP          float64  // Nucleus sampling
    Seed          *int     // Reproducibility
    StopSequences []string // Stop generation
}

Provider-Specific Defaults

Each provider has sensible defaults:

// OpenAI defaults
openai.DefaultProviderDefaults() // temperature: 1.0, max_tokens: 4096

// Claude defaults  
anthropic.DefaultProviderDefaults() // temperature: 1.0, max_tokens: 4096

// Gemini defaults
gemini.DefaultProviderDefaults() // temperature: 0.9, max_tokens: 8192

Implementation Details

OpenAI Provider

Models: GPT-4o, GPT-4o-mini, GPT-4-turbo, GPT-3.5-turbo

Features:

Pricing: Per-token, varies by model

API: REST over HTTPS

Claude Provider (Anthropic)

Models: Claude 3.5 Sonnet, Claude 3.5 Haiku, Claude 3 Opus

Features:

Pricing: Per-token, varies by model

API: REST over HTTPS

Gemini Provider (Google)

Models: Gemini 1.5 Pro, Gemini 1.5 Flash

Features:

Pricing: Per-token, free tier available

API: REST over HTTPS

Message Format

Runtime uses a common message format:

type Message struct {
    Role       string
    Content    string
    ToolCalls  []MessageToolCall
    ToolCallID string
}

Roles:

Translation Layer

Each provider translates to its native format:

Runtime → OpenAI:

{Role: "user", Content: "Hello"}
→
{role: "user", content: "Hello"}

Runtime → Claude:

{Role: "user", Content: "Hello"}
→
{role: "user", content: "Hello"}

Runtime → Gemini:

{Role: "user", Content: "Hello"}
→
{role: "user", parts: [{text: "Hello"}]}

This translation is invisible to users.

Tool Support

All providers support function calling:

Tool Definition

type ToolDef struct {
    Name        string
    Description string
    Parameters  json.RawMessage  // JSON schema
}

Provider-Specific Formats

OpenAI:

{
  "type": "function",
  "function": {
    "name": "get_weather",
    "description": "Get current weather",
    "parameters": { ... }
  }
}

Claude:

{
  "name": "get_weather",
  "description": "Get current weather",
  "input_schema": { ... }
}

Gemini:

{
  "name": "get_weather",
  "description": "Get current weather",
  "parameters": { ... }
}

Runtime handles conversion automatically.

Design Decisions

Why Common Interface?

Decision: All providers implement the same interface

Rationale:

Trade-off: Can’t expose provider-specific features directly. Instead, features are added to the interface when widely supported.

Why Separate Providers?

Decision: One provider instance per LLM service

Rationale:

Alternative considered: Multi-provider registry was considered but rejected as too complex.

Why Not Adapter Pattern?

Decision: Providers translate directly, no adapter layer

Rationale:

Trade-off: Translation code lives in each provider. This is acceptable as translation is straightforward.

Multi-Provider Patterns

Fallback Strategy

Try providers in order:

providers := []types.Provider{primary, secondary, tertiary}

for _, provider := range providers {
    result, err := provider.Complete(ctx, messages, config)
    if err == nil {
        return result, nil
    }
    log.Printf("Provider %s failed: %v", provider.GetProviderName(), err)
}

return nil, errors.New("all providers failed")

Load Balancing

Distribute across providers:

type LoadBalancer struct {
    providers []types.Provider
    current   int
}

func (lb *LoadBalancer) Execute(...) (*ProviderResponse, error) {
    provider := lb.providers[lb.current % len(lb.providers)]
    lb.current++
    return provider.Complete(...)
}

Cost-Based Routing

Route to cheapest provider:

func selectProvider(taskComplexity string) types.Provider {
    switch taskComplexity {
    case "simple":
        return openaiMini  // Cheapest
    case "complex":
        return claude      // Best quality
    case "long_context":
        return gemini      // Largest context
    default:
        return openaiMini
    }
}

Testing with Mock Provider

Runtime includes a mock provider:

mockProvider := mock.NewMockProvider()
mockProvider.SetResponse("Hello! How can I help?")

pipe := pipeline.NewPipeline(
    middleware.ProviderMiddleware(mockProvider, nil, nil, config),
)

result, _ := pipe.Execute(ctx, "user", "Hi")
// result.Response.Content == "Hello! How can I help?"

Benefits:

Performance Considerations

Connection Reuse

Providers maintain HTTP connection pools:

// Good: Reuse provider
provider := openai.NewOpenAIProvider(...)
defer provider.Close()

for _, prompt := range prompts {
    provider.Complete(ctx, messages, config)  // Reuses connections
}

Resource Cleanup

Always close providers:

provider := openai.NewOpenAIProvider(...)
defer provider.Close()  // Essential!

Without closing:

Timeout Management

Use contexts for timeouts:

ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
defer cancel()

result, err := provider.Complete(ctx, messages, config)

Future Extensibility

The provider interface can be extended to support:

New Features:

New Providers:

Summary

Provider abstraction provides:

✅ Vendor Independence: Switch providers easily
✅ Unified API: Same code for all providers
✅ Fallback Support: Try multiple providers
✅ Testing: Mock providers for tests
✅ Extensibility: Add providers without breaking changes

Further Reading