Ga naar hoofdinhoud

Model Selection for Tool-Oriented Agents

Overview​

When creating AI agents that use tools (function calling), model selection is the most important factor for success. Not all LLMs can reliably execute functions - many will describe what they would do instead of actually doing it.

This guide helps you select the right model for tool-oriented agents in OpenRegister.

Quick Recommendations​

For Production Tool Agents​

Best Choice: mistral-nemo (Ollama)

  • ⭐⭐⭐⭐⭐ Excellent function calling
  • Free and runs locally
  • 7.1 GB model size
  • Reliable parameter handling

Alternative: gpt-4o-mini (OpenAI)

  • ⭐⭐⭐⭐⭐ Excellent function calling
  • Requires API key and costs money
  • Best overall accuracy

For Development/Testing​

Good Choice: llama3.2:3b (Ollama)

  • ⭐⭐⭐ Good function calling
  • Lightweight (2 GB)
  • Fast for testing
  • Some inconsistencies

Detailed Model Comparison​

Ollama Models​

⭐⭐⭐⭐⭐ Mistral Nemo (RECOMMENDED)​

Model Name: mistral-nemo

Size: 7.1 GB

Function Calling Quality:

  • ✅ Consistently calls functions when appropriate
  • ✅ Correctly formats all parameter types
  • ✅ Handles complex nested parameters
  • ✅ Low hallucination rate
  • ✅ Good at asking for missing information

Performance:

  • Speed: Fast (even on CPU)
  • Memory: ~8 GB RAM recommended
  • Context: 128K tokens

Use Cases:

  • CMS content management (pages, menus)
  • Object CRUD operations
  • Multi-step workflows
  • Production deployments

Installation:

docker exec openregister-ollama ollama pull mistral-nemo

Agent Configuration:

{
  'provider': 'ollama',
  'model': 'mistral-nemo',
  'temperature': 0.3,
  'tools': ['opencatalogi.cms']
}

⭐⭐⭐⭐ Llama 3.1 8B​

Model Name: llama3.1:8b

Size: 4.7 GB

Function Calling Quality:

  • ✅ Reliable function execution
  • ✅ Good parameter handling
  • ⚠️ Sometimes verbose in responses
  • ✅ Handles follow-up well

Performance:

  • Speed: Fast
  • Memory: ~6 GB RAM
  • Context: 128K tokens

Use Cases:

  • General tool usage
  • Conversational agents with tools
  • Budget-friendly option

⭐⭐⭐ Llama 3.2 3B​

Model Name: llama3.2:3b

Size: 2.0 GB

Function Calling Quality:

  • ✅ Basic function calling works
  • ⚠️ Sometimes describes instead of executing
  • ⚠️ May need explicit prompting
  • ✅ Lightweight and fast

Performance:

  • Speed: Very fast
  • Memory: ~4 GB RAM
  • Context: 128K tokens

Use Cases:

  • Development and testing
  • Simple tool operations
  • Resource-constrained environments

⭐⭐ Mistral 7B (NOT RECOMMENDED FOR TOOLS)​

Model Name: mistral:7b

Size: 4.4 GB

Function Calling Quality:

  • ❌ Unreliable function calling
  • ❌ Often describes what it would do
  • ❌ Inconsistent parameter formatting
  • ⚠️ May work with very explicit prompts

Recommendation: DO NOT USE for tool-oriented agents. Use mistral-nemo instead.

❌ Not Suitable for Function Calling​

These models should NOT be used for agents with tools:

  • qwen2:0.5b - Too small for function calling
  • phi3:mini - Limited function support
  • tinyllama - No function calling capability

OpenAI Models​

⭐⭐⭐⭐⭐ GPT-4o Mini (EXCELLENT)​

Model Name: gpt-4o-mini

Function Calling Quality:

  • ✅ Excellent function execution
  • ✅ Perfect parameter handling
  • ✅ Minimal hallucination
  • ✅ Great at multi-step workflows

Use Cases:

  • Production deployments
  • Complex workflows
  • High-reliability requirements

Cost: Paid API (usage-based pricing)

Installation Guide​

# Connect to your Ollama container
docker exec openregister-ollama ollama pull mistral-nemo

# Verify installation
docker exec openregister-ollama ollama list

Step 2: Configure Agent​

When creating a tool-oriented agent:

  1. Navigate to OpenRegister > Agents
  2. Create new agent
  3. Set Provider: ollama
  4. Set Model: mistral-nemo
  5. Add tools (e.g., opencatalogi.cms)
  6. Set Temperature: 0.3 (lower = more reliable)

Step 3: Test Function Calling​

# Test via API
curl -X POST http://localhost/index.php/apps/openregister/api/chat/send \
  -H 'Content-Type: application/json' \
  -u 'admin:admin' \
  -d '{
    'agentUuid': 'your-agent-uuid',
    'message': 'Create a menu called Test'
  }'

Expected Behavior:

  • ✅ Agent should ASK for required information (items, position)
  • ✅ Once info gathered, should EXECUTE the tool
  • ✅ Should return success message with created object details

Bad Behavior (wrong model):

  • ❌ Agent outputs JSON instead of executing
  • ❌ Agent describes what it would do
  • ❌ Agent ignores tool availability

Troubleshooting​

Issue: Agent describes actions but doesn't execute​

Cause: Model doesn't support function calling well

Solution: Switch to mistral-nemo or llama3.1:8b

Issue: Agent calls functions with wrong parameters​

Cause: Model has poor parameter understanding

Solution:

  1. Switch to mistral-nemo
  2. Lower temperature to 0.2-0.3
  3. Improve system prompt with parameter examples

Issue: Agent never calls functions​

Cause:

  • Model doesn't support tools
  • Tools not properly registered
  • Agent not configured with tools

Solution:

  1. Verify model supports function calling (see table above)
  2. Check tools are registered: See Tool Registration
  3. Verify agent has tools assigned in configuration

Performance Optimization​

For Tool-Oriented Agents​

Optimal Settings:

{
  'provider': 'ollama',
  'model': 'mistral-nemo',
  'temperature': 0.3,
  'maxTokens': 2000,
  'topP': 0.9,
  'frequencyPenalty': 0.0,
  'presencePenalty': 0.0
}

Why these settings?

  • Low temperature (0.3): More deterministic function calling
  • Higher maxTokens (2000): Allows detailed responses after tool execution
  • Standard topP (0.9): Balanced sampling
  • No penalties: Avoid interfering with function formatting

For Chat-Only Agents (No Tools)​

Optimal Settings:

{
  'provider': 'ollama',
  'model': 'llama3.2:3b',
  'temperature': 0.7,
  'maxTokens': 1000
}

Migration Guide​

Upgrading from mistral:7b to mistral-nemo​

  1. Install mistral-nemo:

    docker exec openregister-ollama ollama pull mistral-nemo

  2. Update agent configuration:

    • Navigate to agent settings
    • Change model from mistral:7b to mistral-nemo
    • Lower temperature to 0.3
    • Save configuration

  3. Test thoroughly:

    • Test each tool the agent uses
    • Verify parameter handling
    • Check multi-step workflows

  4. Expected improvements:

    • ✅ Functions actually execute (vs. being described)
    • ✅ Correct parameter formatting
    • ✅ Better conversation flow
    • ✅ Fewer hallucinations

Real-World Examples​

CMS Content Manager Agent​

Configuration:

{
  'name': 'CMS Content Manager',
  'provider': 'ollama',
  'model': 'mistral-nemo',
  'temperature': 0.3,
  'tools': ['opencatalogi.cms'],
  'prompt': 'You are a helpful CMS assistant...'
}

Behavior:

  • Asks for menu title, items, position
  • Executes cms_create_menu function when complete
  • Returns confirmation with created menu details

Document Search Agent​

Configuration:

{
  'name': 'Document Search',
  'provider': 'ollama',
  'model': 'llama3.2:3b',
  'temperature': 0.5,
  'tools': ['openregister.objects'],
  'ragEnabled': true,
  'searchMode': 'hybrid'
}

Behavior:

  • Searches documents using RAG
  • Can execute searches with filters
  • Lower resource requirements

Summary​

For tool-oriented agents, use mistral-nemo. It provides the best balance of:

  • ✅ Reliability
  • ✅ Performance
  • ✅ Cost (free)
  • ✅ Function calling quality

Avoid using mistral:7b, qwen2:0.5b, or phi3:mini for agents that need to execute functions.