managing-apps/assets/documentation/MCP-Configuration-Models.md

# MCP LLM Model Configuration

## Overview

All LLM provider models are now configured exclusively through `appsettings.json` - **no hardcoded values in the code**. This allows you to easily change models without recompiling the application.

## Configuration Location

All model settings are in: `src/Managing.Api/appsettings.json`

```json
{
  "Llm": {
    "Gemini": {
      "ApiKey": "",  // Add your key here or via user secrets
      "DefaultModel": "gemini-3-flash-preview"
    },
    "OpenAI": {
      "ApiKey": "",
      "DefaultModel": "gpt-4o"
    },
    "Claude": {
      "ApiKey": "",
      "DefaultModel": "claude-haiku-4-5-20251001"
    }
  }
}
```

## Current Models (from appsettings.json)

- **Gemini**: `gemini-3-flash-preview`
- **OpenAI**: `gpt-4o`
- **Claude**: `claude-haiku-4-5-20251001`

## Fallback Models (in code)

If `DefaultModel` is not specified in configuration, the providers use these fallback models:

- **Gemini**: `gemini-2.0-flash-exp`
- **OpenAI**: `gpt-4o`
- **Claude**: `claude-3-5-sonnet-20241022`

## How It Works

### 1. Configuration Reading

When the application starts, `LlmService` reads the model configuration:

```csharp
var geminiModel = _configuration["Llm:Gemini:DefaultModel"];
var openaiModel = _configuration["Llm:OpenAI:DefaultModel"];
var claudeModel = _configuration["Llm:Claude:DefaultModel"];
```

### 2. Provider Initialization

Each provider is initialized with the configured model:

```csharp
_providers["gemini"] = new GeminiProvider(geminiApiKey, geminiModel, httpClientFactory, _logger);
_providers["openai"] = new OpenAiProvider(openaiApiKey, openaiModel, httpClientFactory, _logger);
_providers["claude"] = new ClaudeProvider(claudeApiKey, claudeModel, httpClientFactory, _logger);
```

### 3. Model Usage

The provider uses the configured model for all API calls:

```csharp
public async Task<LlmChatResponse> ChatAsync(LlmChatRequest request)
{
    var model = _defaultModel; // From configuration
    var url = $"{BaseUrl}/models/{model}:generateContent?key={_apiKey}";
    // ...
}
```

## Changing Models

### Method 1: Edit appsettings.json

```json
{
  "Llm": {
    "Claude": {
      "DefaultModel": "claude-3-5-sonnet-20241022"  // Change to Sonnet
    }
  }
}
```

### Method 2: Environment Variables

```bash
export Llm__Claude__DefaultModel="claude-3-5-sonnet-20241022"
```

### Method 3: User Secrets (Development)

```bash
cd src/Managing.Api
dotnet user-secrets set "Llm:Claude:DefaultModel" "claude-3-5-sonnet-20241022"
```

## Available Models

### Gemini Models

- `gemini-2.0-flash-exp` - Latest Flash (experimental)
- `gemini-3-flash-preview` - Flash preview
- `gemini-1.5-pro` - Pro model
- `gemini-1.5-flash` - Fast and efficient

### OpenAI Models

- `gpt-4o` - GPT-4 Optimized (recommended)
- `gpt-4o-mini` - Smaller, faster
- `gpt-4-turbo` - GPT-4 Turbo
- `gpt-3.5-turbo` - Cheaper, faster

### Claude Models

- `claude-haiku-4-5-20251001` - Haiku 4.5 (fastest, cheapest)
- `claude-3-5-sonnet-20241022` - Sonnet 3.5 (balanced, recommended)
- `claude-3-opus-20240229` - Opus (most capable)
- `claude-3-sonnet-20240229` - Sonnet 3
- `claude-3-haiku-20240307` - Haiku 3

## Model Selection Guide

### For Development/Testing
- **Gemini**: `gemini-2.0-flash-exp` (free tier)
- **Claude**: `claude-haiku-4-5-20251001` (cheapest)
- **OpenAI**: `gpt-4o-mini` (cheapest)

### For Production (Balanced)
- **Claude**: `claude-3-5-sonnet-20241022` ✅ Recommended
- **OpenAI**: `gpt-4o`
- **Gemini**: `gemini-1.5-pro`

### For Maximum Capability
- **Claude**: `claude-3-opus-20240229` (best reasoning)
- **OpenAI**: `gpt-4-turbo`
- **Gemini**: `gemini-1.5-pro`

### For Speed/Cost Efficiency
- **Claude**: `claude-haiku-4-5-20251001`
- **OpenAI**: `gpt-4o-mini`
- **Gemini**: `gemini-2.0-flash-exp`

## Cost Comparison (Approximate)

### Claude
- **Haiku 4.5**: ~$0.50 per 1M tokens (cheapest)
- **Sonnet 3.5**: ~$9 per 1M tokens (recommended)
- **Opus**: ~$45 per 1M tokens (most expensive)

### OpenAI
- **GPT-4o-mini**: ~$0.30 per 1M tokens
- **GPT-4o**: ~$10 per 1M tokens
- **GPT-4-turbo**: ~$30 per 1M tokens

### Gemini
- **Free tier**: 15 requests/minute (development)
- **Paid**: ~$0.50 per 1M tokens

## Logging

When providers are initialized, you'll see log messages indicating which model is being used:

```
[Information] Gemini provider initialized with model: gemini-3-flash-preview
[Information] OpenAI provider initialized with model: gpt-4o
[Information] Claude provider initialized with model: claude-haiku-4-5-20251001
```

If no model is configured, it will show:

```
[Information] Gemini provider initialized with model: default
```

And the fallback model will be used.

## Best Practices

1. **Use environment variables** for production to keep configuration flexible
2. **Test with cheaper models** during development
3. **Monitor costs** in provider dashboards
4. **Update models** as new versions are released
5. **Document changes** when switching models for your team

## Example Configurations

### Development (Cost-Optimized)
```json
{
  "Llm": {
    "Claude": {
      "ApiKey": "your-key",
      "DefaultModel": "claude-haiku-4-5-20251001"
    }
  }
}
```

### Production (Balanced)
```json
{
  "Llm": {
    "Claude": {
      "ApiKey": "your-key",
      "DefaultModel": "claude-3-5-sonnet-20241022"
    }
  }
}
```

### High-Performance (Maximum Capability)
```json
{
  "Llm": {
    "Claude": {
      "ApiKey": "your-key",
      "DefaultModel": "claude-3-opus-20240229"
    }
  }
}
```

## Verification

To verify which model is being used:

1. Check application logs on startup
2. Look for provider initialization messages
3. Check LLM response metadata (includes model name)
4. Monitor provider dashboards for API usage

## Troubleshooting

### Model not found error

**Issue**: "Model not found" or "Invalid model name"

**Solution**:
1. Verify model name spelling in `appsettings.json`
2. Check provider documentation for available models
3. Ensure model is available in your region/tier
4. Try removing `DefaultModel` to use the fallback

### Wrong model being used

**Issue**: Application uses fallback instead of configured model

**Solution**:
1. Check configuration path: `Llm:ProviderName:DefaultModel`
2. Verify no typos in JSON (case-sensitive)
3. Restart application after configuration changes
4. Check logs for which model was loaded

### Configuration not loading

**Issue**: Changes to `appsettings.json` not taking effect

**Solution**:
1. Restart the application
2. Clear build artifacts: `dotnet clean`
3. Check file is in correct location: `src/Managing.Api/appsettings.json`
4. Verify JSON syntax is valid

## Summary

✅ All models configured in `appsettings.json`
✅ No hardcoded model names in code
✅ Easy to change without recompiling
✅ Fallback models in case of missing configuration
✅ Full flexibility for different environments
✅ Logged on startup for verification

This design allows maximum flexibility while maintaining sensible defaults!