McFlow

# LLM/AI Model Parameter Validation ## Overview McFlow MCP Server now validates LLM/AI model parameters to ensure they match the exact requirements of each model provider. This prevents runtime errors and ensures optimal model performance. ## ⚠️ CRITICAL: Model-Specific Parameters Each AI provider has DIFFERENT parameter names and requirements. Using the wrong parameters will cause your workflow to fail. ## Validated Providers ### 🟢 OpenAI (GPT-4, GPT-3.5) **Valid Models:** - `gpt-4`, `gpt-4-turbo`, `gpt-4o`, `gpt-4o-mini` - `gpt-3.5-turbo`, `gpt-3.5-turbo-16k` **Correct Parameters:** ```json { "model": "gpt-4", "temperature": 0.7, // 0-2 "max_tokens": 1024, // 1-128000 for GPT-4 "top_p": 0.9, // 0-1 "frequency_penalty": 0, // -2 to 2 "presence_penalty": 0, // -2 to 2 "stop": ["\\n\\n"], // Array of stop sequences "n": 1 // Number of completions } ``` **Common Mistakes:** - ❌ Using `max_length` instead of `max_tokens` - ❌ Using `repetition_penalty` (that's for other models) - ❌ Using deprecated models like `text-davinci-003` ### 🔵 Anthropic (Claude) **Valid Models:** - `claude-3-opus-20240229` - `claude-3-sonnet-20240229` - `claude-3-5-sonnet-20241022` - `claude-3-haiku-20240307` **Correct Parameters:** ```json { "model": "claude-3-opus-20240229", "max_tokens": 1024, // REQUIRED! 1-4096 "temperature": 0.7, // 0-1 (not 0-2 like OpenAI!) "top_p": 0.9, // 0-1 "top_k": 40, // Claude-specific "stop_sequences": ["\\n"] // Note: stop_sequences, not stop } ``` **Common Mistakes:** - ❌ Forgetting `max_tokens` (REQUIRED for Claude!) - ❌ Using `frequency_penalty` or `presence_penalty` (OpenAI-only) - ❌ Temperature > 1 (Claude max is 1, not 2) ### 🔴 Google AI (Gemini) **Valid Models:** - `gemini-pro` - `gemini-1.5-pro` - `gemini-1.5-flash` **Correct Parameters:** ```json { "model": "gemini-pro", "maxOutputTokens": 2048, // NOT max_tokens! "temperature": 0.7, // 0-1 "topP": 0.9, // camelCase! "topK": 40, // camelCase! "candidateCount": 1 // Must be 1 currently } ``` **Common Mistakes:** - ❌ Using `max_tokens` instead of `maxOutputTokens` - ❌ Using snake_case instead of camelCase - ❌ Setting `candidateCount` > 1 ### 🟡 Cohere **Correct Parameters:** ```json { "model": "command-r", "max_tokens": 1024, "temperature": 0.7, // 0-5 (higher than others!) "p": 0.9, // NOT top_p! "k": 40, // NOT top_k! "frequency_penalty": 0, "presence_penalty": 0 } ``` **Common Mistakes:** - ❌ Using `top_p` instead of `p` - ❌ Using `top_k` instead of `k` ### 🟣 Replicate (Llama, Mistral, etc.) **Correct Parameters:** ```json { "model": "meta/llama-2-70b-chat", "max_new_tokens": 1024, // NOT max_tokens! "temperature": 0.7, // Must be > 0 (min 0.01) "top_p": 0.9, "repetition_penalty": 1.1 // Replicate uses this } ``` **Common Mistakes:** - ❌ Using `max_tokens` instead of `max_new_tokens` - ❌ Setting temperature to 0 (min is 0.01) ## Validation Examples ### ❌ Invalid OpenAI Configuration ```json { "type": "n8n-nodes-base.openAi", "parameters": { "model": "text-davinci-003", // Deprecated! "max_length": 1024, // Wrong parameter name "repetition_penalty": 1.1 // Not an OpenAI parameter } } ``` **Validation Errors:** - `model: text-davinci models are deprecated → Use gpt-3.5-turbo or gpt-4` - `max_length: Invalid parameter → Use 'max_tokens' instead` - `repetition_penalty: Invalid for OpenAI → Use 'frequency_penalty' instead` ### ✅ Fixed OpenAI Configuration ```json { "type": "n8n-nodes-base.openAi", "parameters": { "model": "gpt-4", "max_tokens": 1024, "frequency_penalty": 0.1 } } ``` ## Auto-Fix Capabilities McFlow can automatically fix common LLM parameter issues: 1. **Parameter Name Corrections:** - OpenAI: `maxTokens` → `max_tokens` - Google: `max_tokens` → `maxOutputTokens` - Cohere: `top_p` → `p` - Replicate: `max_tokens` → `max_new_tokens` 2. **Model Updates:** - `text-davinci-*` → `gpt-3.5-turbo` - `palm-*` → `gemini-pro` 3. **Required Parameters:** - Adds `max_tokens` for Claude if missing - Sets `candidateCount: 1` for Google 4. **Invalid Values:** - Replicate `temperature: 0` → `0.01` - Removes OpenAI-specific params from Claude ## Validation Commands ```bash # Validate all workflows for LLM parameter issues McFlow validate # Auto-fix LLM parameter issues McFlow validate --fix # Validate specific workflow McFlow validate --workflow my-ai-workflow ``` ## Best Practices 1. **Always specify the model explicitly** - Don't rely on defaults 2. **Check parameter limits** - Each model has different ranges 3. **Use correct parameter names** - They vary by provider 4. **Include required parameters** - Some are mandatory (e.g., Claude's max_tokens) 5. **Test with small values first** - Start with lower token limits 6. **Check the provider's docs** - APIs change frequently ## Common Cross-Provider Differences | Parameter | OpenAI | Claude | Google | Cohere | Replicate | |-----------|--------|--------|--------|--------|-----------| | Max Tokens | `max_tokens` | `max_tokens` ✅ | `maxOutputTokens` | `max_tokens` | `max_new_tokens` | | Temperature Range | 0-2 | 0-1 | 0-1 | 0-5 | 0.01-5 | | Top P | `top_p` | `top_p` | `topP` | `p` | `top_p` | | Top K | ❌ | `top_k` | `topK` | `k` | `top_k` | | Stop | `stop` | `stop_sequences` | `stopSequences` | `stop_sequences` | `stop` | | Required | None | `max_tokens` | None | None | None | ## Error Messages When validation fails, you'll see specific error messages: ``` ❌ LLM Parameter Error - temperature: Invalid temperature: 3. Must be between 0 and 1 → Fix: Set temperature between 0 and 1 ❌ LLM Parameter Error - max_tokens: Claude models require max_tokens parameter → Fix: Add max_tokens parameter (e.g., 1024) ❌ LLM Parameter Error - model: text-davinci models are deprecated → Fix: Use gpt-3.5-turbo or gpt-4 instead ``` ## Getting Help - **OpenAI Docs**: https://platform.openai.com/docs/api-reference - **Anthropic Docs**: https://docs.anthropic.com/claude/reference - **Google AI Docs**: https://ai.google.dev/api/rest - **Cohere Docs**: https://docs.cohere.com/reference - **Replicate Docs**: https://replicate.com/docs Remember: Each provider's API is different. What works for OpenAI might not work for Claude!