Merge pull request #819 from MervinPraison/claude/issue-818-20250711_091438

fix: enable tool calling for Gemini models

Author: MervinPraison

src/praisonai-agents/praisonaiagents/agent/agent.py

@@ -713,26 +713,69 @@ def _apply_guardrail_with_retry(self, response_text, prompt, temperature=0.2, to
                 )
 
         return current_response
+    
+    def _build_system_prompt(self, tools=None):
+        """Build the system prompt with tool information.
+        
+        Args:
+            tools: Optional list of tools to use (defaults to self.tools)
+            
+        Returns:
+            str: The system prompt or None if use_system_prompt is False
+        """
+        if not self.use_system_prompt:
+            return None
+            
+        system_prompt = f"""{self.backstory}\n
+Your Role: {self.role}\n
+Your Goal: {self.goal}"""
+        
+        # Add tool usage instructions if tools are available
+        # Use provided tools or fall back to self.tools
+        tools_to_use = tools if tools is not None else self.tools
+        if tools_to_use:
+            tool_names = []
+            for tool in tools_to_use:
+                try:
+                    if callable(tool) and hasattr(tool, '__name__'):
+                        tool_names.append(tool.__name__)
+                    elif isinstance(tool, dict) and isinstance(tool.get('function'), dict) and 'name' in tool['function']:
+                        tool_names.append(tool['function']['name'])
+                    elif isinstance(tool, str):
+                        tool_names.append(tool)
+                    elif hasattr(tool, "to_openai_tool"):
+                        # Handle MCP tools
+                        openai_tools = tool.to_openai_tool()
+                        if isinstance(openai_tools, list):
+                            for t in openai_tools:
+                                if isinstance(t, dict) and 'function' in t and 'name' in t['function']:
+                                    tool_names.append(t['function']['name'])
+                        elif isinstance(openai_tools, dict) and 'function' in openai_tools:
+                            tool_names.append(openai_tools['function']['name'])
+                except (AttributeError, KeyError, TypeError) as e:
+                    logging.warning(f"Could not extract tool name from {tool}: {e}")
+                    continue
+            
+            if tool_names:
+                system_prompt += f"\n\nYou have access to the following tools: {', '.join(tool_names)}. Use these tools when appropriate to help complete your tasks. Always use tools when they can help provide accurate information or perform actions."
+        
+        return system_prompt
 
-    def _build_messages(self, prompt, temperature=0.2, output_json=None, output_pydantic=None):
+    def _build_messages(self, prompt, temperature=0.2, output_json=None, output_pydantic=None, tools=None):
         """Build messages list for chat completion.
         
         Args:
             prompt: The user prompt (str or list)
             temperature: Temperature for the chat
             output_json: Optional Pydantic model for JSON output
             output_pydantic: Optional Pydantic model for JSON output (alias)
+            tools: Optional list of tools to use (defaults to self.tools)
             
         Returns:
             tuple: (messages list, original prompt)
         """
-        # Build system prompt if enabled
-        system_prompt = None
-        if self.use_system_prompt:
-            system_prompt = f"""{self.backstory}\n
-Your Role: {self.role}\n
-Your Goal: {self.goal}
-            """
+        # Build system prompt using the helper method
+        system_prompt = self._build_system_prompt(tools)
 
         # Use openai_client's build_messages method if available
         if self._openai_client is not None:
@@ -1176,7 +1219,7 @@ def chat(self, prompt, temperature=0.2, tools=None, output_json=None, output_pyd
                     # Pass everything to LLM class
                     response_text = self.llm_instance.get_response(
                     prompt=prompt,
-                    system_prompt=f"{self.backstory}\n\nYour Role: {self.role}\n\nYour Goal: {self.goal}" if self.use_system_prompt else None,
+                    system_prompt=self._build_system_prompt(tools),
                     chat_history=self.chat_history,
                     temperature=temperature,
                     tools=tool_param,
@@ -1492,7 +1535,7 @@ async def achat(self, prompt: str, temperature=0.2, tools=None, output_json=None
                 try:
                     response_text = await self.llm_instance.get_response_async(
                         prompt=prompt,
-                        system_prompt=f"{self.backstory}\n\nYour Role: {self.role}\n\nYour Goal: {self.goal}" if self.use_system_prompt else None,
+                        system_prompt=self._build_system_prompt(tools),
                         chat_history=self.chat_history,
                         temperature=temperature,
                         tools=tools,
@@ -1506,7 +1549,7 @@ async def achat(self, prompt: str, temperature=0.2, tools=None, output_json=None
                         console=self.console,
                         agent_name=self.name,
                         agent_role=self.role,
-                        agent_tools=[t.__name__ if hasattr(t, '__name__') else str(t) for t in self.tools],
+                        agent_tools=[t.__name__ if hasattr(t, '__name__') else str(t) for t in (tools if tools is not None else self.tools)],
                         execute_tool_fn=self.execute_tool_async,
                         reasoning_steps=reasoning_steps
                     )

src/praisonai-agents/praisonaiagents/llm/llm.py

@@ -406,7 +406,7 @@ def _supports_streaming_tools(self) -> bool:
         # missing tool calls or making duplicate calls
         return False
 
-    def _build_messages(self, prompt, system_prompt=None, chat_history=None, output_json=None, output_pydantic=None):
+    def _build_messages(self, prompt, system_prompt=None, chat_history=None, output_json=None, output_pydantic=None, tools=None):
         """Build messages list for LLM completion. Works for both sync and async.
         
         Args:
@@ -415,6 +415,7 @@ def _build_messages(self, prompt, system_prompt=None, chat_history=None, output_
             chat_history: Optional list of previous messages
             output_json: Optional Pydantic model for JSON output
             output_pydantic: Optional Pydantic model for JSON output (alias)
+            tools: Optional list of tools available
             
         Returns:
             tuple: (messages list, original prompt)
@@ -1858,6 +1859,21 @@ def _build_completion_params(self, **override_params) -> Dict[str, Any]:
         # Override with any provided parameters
         params.update(override_params)
 
+        # Add tool_choice="auto" when tools are provided (unless already specified)
+        if 'tools' in params and params['tools'] and 'tool_choice' not in params:
+            # For Gemini models, use tool_choice to encourage tool usage
+            # More comprehensive Gemini model detection
+            if any(prefix in self.model.lower() for prefix in ['gemini', 'gemini/', 'google/gemini']):
+                try:
+                    import litellm
+                    # Check if model supports function calling before setting tool_choice
+                    if litellm.supports_function_calling(model=self.model):
+                        params['tool_choice'] = 'auto'
+                except Exception as e:
+                    # If check fails, still set tool_choice for known Gemini models
+                    logging.debug(f"Could not verify function calling support: {e}. Setting tool_choice anyway.")
+                    params['tool_choice'] = 'auto'
+        
         return params
 
     def _prepare_response_logging(self, temperature: float, stream: bool, verbose: bool, markdown: bool, **kwargs) -> Optional[Dict[str, Any]]:

src/praisonai/tests/unit/test_async_gemini_fix.py

@@ -0,0 +1,163 @@
+"""
+Test to verify that async agents with Gemini models properly use tools
+after the fix for issue #818
+"""
+import asyncio
+import logging
+from praisonaiagents import Agent, Task, PraisonAIAgents
+
+# Enable logging to see tool calls
+logging.basicConfig(level=logging.INFO)
+
+# Mock search tool
+async def mock_search(query: str) -> dict:
+    """Mock search tool for testing"""
+    print(f"[TOOL CALLED] Searching for: {query}")
+    return {
+        "query": query,
+        "results": [
+            {
+                "title": f"Result 1 for {query}",
+                "snippet": f"This is a mock result about {query}",
+                "url": "https://example.com/1"
+            },
+            {
+                "title": f"Result 2 for {query}", 
+                "snippet": f"Another mock result about {query}",
+                "url": "https://example.com/2"
+            }
+        ],
+        "status": "success"
+    }
+
+async def test_async_gemini_tools():
+    """Test async agents with Gemini models use tools correctly"""
+    
+    # Create search agent with Gemini model
+    search_agent = Agent(
+        name="AsyncSearcher",
+        role="Research Assistant",
+        goal="Find information using the search tool",
+        backstory="You are an expert at finding information online",
+        tools=[mock_search],
+        llm={"model": "gemini/gemini-1.5-flash-latest"},
+        verbose=True
+    )
+    
+    # Create analysis agent without tools
+    analysis_agent = Agent(
+        name="Analyzer",
+        role="Data Analyst",
+        goal="Analyze search results",
+        backstory="You excel at analyzing and summarizing information",
+        llm={"model": "gemini/gemini-1.5-flash-latest"},
+        verbose=True
+    )
+    
+    # Create tasks
+    search_task = Task(
+        name="search_task",
+        description="Search for information about 'quantum computing breakthroughs 2024'",
+        expected_output="Search results with at least 2 relevant findings",
+        agent=search_agent,
+        async_execution=True
+    )
+    
+    analysis_task = Task(
+        name="analysis_task", 
+        description="Analyze the search results and provide a summary",
+        expected_output="A concise summary of the findings",
+        agent=analysis_agent,
+        context=[search_task],
+        async_execution=False
+    )
+    
+    # Create workflow
+    workflow = PraisonAIAgents(
+        agents=[search_agent, analysis_agent],
+        tasks=[search_task, analysis_task],
+        verbose=True
+    )
+    
+    # Execute async
+    print("\n🚀 Starting async agent test with Gemini models...")
+    result = await workflow.astart()
+    
+    # Check results
+    print("\n✅ Test Results:")
+    print("-" * 50)
+    
+    # Verify search agent used the tool
+    search_result = str(result)
+    if "mock result" in search_result.lower() or "tool called" in search_result.lower():
+        print("✅ SUCCESS: Search agent properly used the mock_search tool!")
+    else:
+        print("❌ FAILURE: Search agent did NOT use the tool (claimed no internet access)")
+    
+    # Show the actual output
+    print("\nFinal output:")
+    print(result)
+    
+    return result
+
+async def test_multiple_async_agents():
+    """Test multiple async agents running in parallel"""
+    
+    agents = []
+    tasks = []
+    
+    # Create 3 search agents
+    for i in range(3):
+        agent = Agent(
+            name=f"AsyncAgent{i}",
+            role="Researcher",
+            goal="Search for information",
+            backstory="Expert researcher",
+            tools=[mock_search],
+            llm={"model": "gemini/gemini-1.5-flash-latest"}
+        )
+        
+        task = Task(
+            name=f"task_{i}",
+            description=f"Search for 'AI advancement #{i+1}'",
+            expected_output="Search results",
+            agent=agent,
+            async_execution=True
+        )
+        
+        agents.append(agent)
+        tasks.append(task)
+    
+    # Execute all in parallel
+    workflow = PraisonAIAgents(agents=agents, tasks=tasks)
+    
+    print("\n🚀 Testing multiple async agents in parallel...")
+    results = await workflow.astart()
+    
+    # Verify all agents used tools
+    success_count = 0
+    for i, task in enumerate(tasks):
+        if "mock result" in str(results).lower():
+            success_count += 1
+    
+    print(f"\n✅ {success_count}/{len(tasks)} agents successfully used tools")
+    
+    return results
+
+async def main():
+    """Run all async tests"""
+    try:
+        # Test 1: Single async agent
+        await test_async_gemini_tools()
+        
+        # Test 2: Multiple async agents in parallel
+        await test_multiple_async_agents()
+        
+        print("\n🎉 All async tests completed!")
+        
+    except Exception as e:
+        print(f"\n❌ Error during testing: {e}")
+        raise
+
+if __name__ == "__main__":
+    asyncio.run(main())

src/praisonai/tests/unit/test_gemini_tool_choice.py

@@ -0,0 +1,64 @@
+"""
+Simple test to verify that tool_choice='auto' is set for Gemini models
+"""
+import logging
+from praisonaiagents.llm.llm import LLM
+
+# Enable debug logging to see our log message
+logging.basicConfig(level=logging.DEBUG, format='%(levelname)s: %(message)s')
+
+# Test different Gemini model formats
+test_models = [
+    "gemini/gemini-1.5-flash-8b",
+    "gemini-1.5-flash-8b",
+    "gemini/gemini-pro",
+    "gpt-4",  # Non-Gemini model for comparison
+]
+
+# Mock tools
+mock_tools = [
+    {
+        "type": "function",
+        "function": {
+            "name": "search",
+            "description": "Search for information",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "query": {"type": "string"}
+                }
+            }
+        }
+    }
+]
+
+print("Testing tool_choice setting for different models:\n")
+
+for model in test_models:
+    print(f"\nTesting model: {model}")
+    try:
+        llm = LLM(model=model)
+        params = llm._build_completion_params(
+            messages=[{"role": "user", "content": "test"}],
+            tools=mock_tools
+        )
+        
+        tool_choice = params.get('tool_choice', 'NOT SET')
+        print(f"  tool_choice: {tool_choice}")
+        
+        # Verify behavior
+        if model.startswith(('gemini-', 'gemini/')):
+            if tool_choice == 'auto':
+                print(f"  ✅ CORRECT: Gemini model has tool_choice='auto'")
+            else:
+                print(f"  ❌ ERROR: Gemini model should have tool_choice='auto'")
+        else:
+            if tool_choice == 'NOT SET':
+                print(f"  ✅ CORRECT: Non-Gemini model doesn't have tool_choice set")
+            else:
+                print(f"  ⚠️  WARNING: Non-Gemini model has tool_choice set to '{tool_choice}'")
+                
+    except Exception as e:
+        print(f"  ❌ ERROR: {e}")
+
+print("\nTest complete!")