praisonaiagents 0.0.118__tar.gz → 0.0.123__tar.gz

{praisonaiagents-0.0.118 → praisonaiagents-0.0.123}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: praisonaiagents
-Version: 0.0.118
+Version: 0.0.123
 Summary: Praison AI agents for completing complex tasks with Self Reflection Agents
 Author: Mervin Praison
 Requires-Python: >=3.10

{praisonaiagents-0.0.118 → praisonaiagents-0.0.123}/praisonaiagents/agent/agent.py

@@ -1233,7 +1233,8 @@ Your Goal: {self.goal}
                         # Add to chat history and return raw response
                         self.chat_history.append({"role": "user", "content": original_prompt})
                         self.chat_history.append({"role": "assistant", "content": response_text})
-                        if self.verbose:
+                        # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
+                        if self.verbose and not self._using_custom_llm:
                             display_interaction(original_prompt, response_text, markdown=self.markdown, 
                                              generation_time=time.time() - start_time, console=self.console)
                         return response_text
@@ -1243,7 +1244,9 @@ Your Goal: {self.goal}
                         self.chat_history.append({"role": "assistant", "content": response_text})
                         if self.verbose:
                             logging.debug(f"Agent {self.name} final response: {response_text}")
-                        display_interaction(original_prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
+                        # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
+                        if self.verbose and not self._using_custom_llm:
+                            display_interaction(original_prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
                         # Return only reasoning content if reasoning_steps is True
                         if reasoning_steps and hasattr(response.choices[0].message, 'reasoning_content'):
                             # Apply guardrail to reasoning content
@@ -1271,25 +1274,40 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     messages.append({"role": "user", "content": reflection_prompt})
 
                     try:
-                        # Check if OpenAI client is available
-                        if self._openai_client is None:
-                            # For custom LLMs, self-reflection with structured output is not supported
-                            if self.verbose:
-                                display_self_reflection(f"Agent {self.name}: Self-reflection with structured output is not supported for custom LLM providers. Skipping reflection.", console=self.console)
-                            # Return the original response without reflection
-                            self.chat_history.append({"role": "user", "content": prompt})
-                            self.chat_history.append({"role": "assistant", "content": response_text})
-                            display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
-                            return response_text
-                        
-                        reflection_response = self._openai_client.sync_client.beta.chat.completions.parse(
-                            model=self.reflect_llm if self.reflect_llm else self.llm,
-                            messages=messages,
-                            temperature=temperature,
-                            response_format=ReflectionOutput
-                        )
+                        # Check if we're using a custom LLM (like Gemini)
+                        if self._using_custom_llm or self._openai_client is None:
+                            # For custom LLMs, we need to handle reflection differently
+                            # Use non-streaming to get complete JSON response
+                            reflection_response = self._chat_completion(messages, temperature=temperature, tools=None, stream=False, reasoning_steps=False)
+                            
+                            if not reflection_response or not reflection_response.choices:
+                                raise Exception("No response from reflection request")
+                            
+                            reflection_text = reflection_response.choices[0].message.content.strip()
+                            
+                            # Clean the JSON output
+                            cleaned_json = self.clean_json_output(reflection_text)
+                            
+                            # Parse the JSON manually
+                            reflection_data = json.loads(cleaned_json)
+                            
+                            # Create a reflection output object manually
+                            class CustomReflectionOutput:
+                                def __init__(self, data):
+                                    self.reflection = data.get('reflection', '')
+                                    self.satisfactory = data.get('satisfactory', 'no').lower()
+                            
+                            reflection_output = CustomReflectionOutput(reflection_data)
+                        else:
+                            # Use OpenAI's structured output for OpenAI models
+                            reflection_response = self._openai_client.sync_client.beta.chat.completions.parse(
+                                model=self.reflect_llm if self.reflect_llm else self.llm,
+                                messages=messages,
+                                temperature=temperature,
+                                response_format=ReflectionOutput
+                            )
 
-                        reflection_output = reflection_response.choices[0].message.parsed
+                            reflection_output = reflection_response.choices[0].message.parsed
 
                         if self.verbose:
                             display_self_reflection(f"Agent {self.name} self reflection (using {self.reflect_llm if self.reflect_llm else self.llm}): reflection='{reflection_output.reflection}' satisfactory='{reflection_output.satisfactory}'", console=self.console)
@@ -1302,7 +1320,9 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 display_self_reflection("Agent marked the response as satisfactory after meeting minimum reflections", console=self.console)
                             self.chat_history.append({"role": "user", "content": prompt})
                             self.chat_history.append({"role": "assistant", "content": response_text})
-                            display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
+                            # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
+                            if self.verbose and not self._using_custom_llm:
+                                display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
                             # Apply guardrail validation after satisfactory reflection
                             try:
                                 validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
@@ -1317,7 +1337,9 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 display_self_reflection("Maximum reflection count reached, returning current response", console=self.console)
                             self.chat_history.append({"role": "user", "content": prompt})
                             self.chat_history.append({"role": "assistant", "content": response_text})
-                            display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
+                            # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
+                            if self.verbose and not self._using_custom_llm:
+                                display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
                             # Apply guardrail validation after max reflections
                             try:
                                 validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
@@ -1328,7 +1350,9 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
 
                         logging.debug(f"{self.name} reflection count {reflection_count + 1}, continuing reflection process")
                         messages.append({"role": "user", "content": "Now regenerate your response using the reflection you made"})
-                        response = self._chat_completion(messages, temperature=temperature, tools=None, stream=self.stream)
+                        # For custom LLMs during reflection, always use non-streaming to ensure complete responses
+                        use_stream = self.stream if not self._using_custom_llm else False
+                        response = self._chat_completion(messages, temperature=temperature, tools=None, stream=use_stream)
                         response_text = response.choices[0].message.content.strip()
                         reflection_count += 1
                         continue  # Continue the loop for more reflections

{praisonaiagents-0.0.118 → praisonaiagents-0.0.123}/praisonaiagents/agents/autoagents.py

@@ -13,7 +13,8 @@ import logging
 import os
 from pydantic import BaseModel, ConfigDict
 from ..main import display_instruction, display_tool_call, display_interaction
-from ..llm import get_openai_client
+from ..llm import get_openai_client, LLM, OpenAIClient
+import json
 
 # Define Pydantic models for structured output
 class TaskConfig(BaseModel):
@@ -108,6 +109,8 @@ class AutoAgents(PraisonAIAgents):
         self.max_execution_time = max_execution_time
         self.max_iter = max_iter
         self.reflect_llm = reflect_llm
+        self.base_url = base_url
+        self.api_key = api_key
         
         # Display initial instruction
         if self.verbose:
@@ -211,9 +214,37 @@ Tools: {', '.join(agent_tools)}"""
             
         return assigned_tools
 
+    def _validate_config(self, config: AutoAgentsConfig) -> tuple[bool, str]:
+        """
+        Validate that the configuration has proper TaskConfig objects.
+        
+        Returns:
+            Tuple of (is_valid, error_message)
+        """
+        for agent_idx, agent in enumerate(config.agents):
+            if not hasattr(agent, 'tasks') or not agent.tasks:
+                return False, f"Agent '{agent.name}' has no tasks defined"
+            
+            for task_idx, task in enumerate(agent.tasks):
+                # Check if task is a proper TaskConfig instance
+                if not isinstance(task, TaskConfig):
+                    return False, f"Task at index {task_idx} for agent '{agent.name}' is not a proper TaskConfig object"
+                
+                # Check required fields
+                if not task.name:
+                    return False, f"Task at index {task_idx} for agent '{agent.name}' has no name"
+                if not task.description:
+                    return False, f"Task at index {task_idx} for agent '{agent.name}' has no description"
+                if not task.expected_output:
+                    return False, f"Task at index {task_idx} for agent '{agent.name}' has no expected_output"
+                if task.tools is None:
+                    return False, f"Task at index {task_idx} for agent '{agent.name}' has no tools field"
+        
+        return True, ""
+
     def _generate_config(self) -> AutoAgentsConfig:
-        """Generate the configuration for agents and tasks"""
-        prompt = f"""
+        """Generate the configuration for agents and tasks with retry logic"""
+        base_prompt = f"""
 Generate a configuration for AI agents to accomplish this task: "{self.instructions}"
 
 The configuration should include:
@@ -234,40 +265,161 @@ Requirements:
 4. The process type should match the task requirements
 5. Generate maximum {self.max_agents} agents to handle this task efficiently
 
-Return the configuration in a structured JSON format matching the AutoAgentsConfig schema.
+Return the configuration in a structured JSON format matching this exact schema:
+{{
+  "main_instruction": "Overall goal description",
+  "process_type": "sequential|workflow|hierarchical",
+  "agents": [
+    {{
+      "name": "Agent Name",
+      "role": "Agent Role",
+      "goal": "Agent Goal",
+      "backstory": "Agent Backstory",
+      "tools": ["tool1", "tool2"],
+      "tasks": [
+        {{
+          "name": "Task Name",
+          "description": "Detailed task description",
+          "expected_output": "What the task should produce",
+          "tools": ["tool1", "tool2"]
+        }}
+      ]
+    }}
+  ]
+}}
+
+IMPORTANT: Each task MUST be an object with name, description, expected_output, and tools fields, NOT a simple string.
 """
         
-        try:
-            # Get OpenAI client
-            try:
-                client = get_openai_client()
-            except ValueError as e:
-                # AutoAgents requires OpenAI for structured output generation
-                raise ValueError(
-                    "AutoAgents requires OpenAI API for automatic agent generation. "
-                    "Please set OPENAI_API_KEY environment variable or use PraisonAIAgents class directly "
-                    "with manually configured agents for non-OpenAI providers."
-                ) from e
-                
-            response = client.beta.chat.completions.parse(
-                model=self.llm,
-                response_format=AutoAgentsConfig,
-                messages=[
-                    {"role": "system", "content": "You are a helpful assistant designed to generate AI agent configurations."},
-                    {"role": "user", "content": prompt}
-                ]
-            )
+        max_retries = 3
+        last_response = None
+        last_error = None
+        
+        for attempt in range(max_retries):
+            # Initialize variables for this attempt
+            use_openai_structured = False
+            client = None
+            
+            # Prepare prompt for this attempt
+            if attempt > 0 and last_response and last_error:
+                # On retry, include the previous response and error
+                prompt = f"""{base_prompt}
+
+PREVIOUS ATTEMPT FAILED!
+Your previous response was:
+```json
+{last_response}
+```
+
+Error: {last_error}
+
+REMEMBER: Tasks MUST be objects with the following structure:
+{{
+  "name": "Task Name",
+  "description": "Task Description",
+  "expected_output": "Expected Output",
+  "tools": ["tool1", "tool2"]
+}}
+
+DO NOT use strings for tasks. Each task MUST be a complete object with all four fields."""
+            else:
+                prompt = base_prompt
             
-            # Ensure we have exactly max_agents number of agents
-            if len(response.choices[0].message.parsed.agents) > self.max_agents:
-                response.choices[0].message.parsed.agents = response.choices[0].message.parsed.agents[:self.max_agents]
-            elif len(response.choices[0].message.parsed.agents) < self.max_agents:
-                logging.warning(f"Generated {len(response.choices[0].message.parsed.agents)} agents, expected {self.max_agents}")
+            try:
+                # Check if we have OpenAI API and the model supports structured output
+                from ..llm import supports_structured_outputs
+                if self.llm and supports_structured_outputs(self.llm):
+                    client = get_openai_client()
+                    use_openai_structured = True
+            except:
+                # If OpenAI client is not available, we'll use the LLM class
+                pass
             
-            return response.choices[0].message.parsed
-        except Exception as e:
-            logging.error(f"Error generating configuration: {e}")
-            raise
+            try:
+                if use_openai_structured and client:
+                    # Use OpenAI's structured output for OpenAI models (backward compatibility)
+                    config = client.parse_structured_output(
+                        messages=[
+                            {"role": "system", "content": "You are a helpful assistant designed to generate AI agent configurations."},
+                            {"role": "user", "content": prompt}
+                        ],
+                        response_format=AutoAgentsConfig,
+                        model=self.llm
+                    )
+                    # Store the response for potential retry
+                    last_response = json.dumps(config.model_dump(), indent=2)
+                else:
+                    # Use LLM class for all other providers (Gemini, Anthropic, etc.)
+                    llm_instance = LLM(
+                        model=self.llm,
+                        base_url=self.base_url,
+                        api_key=self.api_key
+                    )
+                    
+                    response_text = llm_instance.get_response(
+                        prompt=prompt,
+                        system_prompt="You are a helpful assistant designed to generate AI agent configurations.",
+                        output_pydantic=AutoAgentsConfig,
+                        temperature=0.7,
+                        stream=False,
+                        verbose=False
+                    )
+                    
+                    # Store the raw response for potential retry
+                    last_response = response_text
+                    
+                    # Parse the JSON response
+                    try:
+                        # First try to parse as is
+                        config_dict = json.loads(response_text)
+                        config = AutoAgentsConfig(**config_dict)
+                    except json.JSONDecodeError:
+                        # If that fails, try to extract JSON from the response
+                        # Handle cases where the model might wrap JSON in markdown blocks
+                        cleaned_response = response_text.strip()
+                        if cleaned_response.startswith("```json"):
+                            cleaned_response = cleaned_response[7:]
+                        if cleaned_response.startswith("```"):
+                            cleaned_response = cleaned_response[3:]
+                        if cleaned_response.endswith("```"):
+                            cleaned_response = cleaned_response[:-3]
+                        cleaned_response = cleaned_response.strip()
+                        
+                        config_dict = json.loads(cleaned_response)
+                        config = AutoAgentsConfig(**config_dict)
+                
+                # Validate the configuration
+                is_valid, error_msg = self._validate_config(config)
+                if not is_valid:
+                    last_error = error_msg
+                    if attempt < max_retries - 1:
+                        logging.warning(f"Configuration validation failed (attempt {attempt + 1}/{max_retries}): {error_msg}")
+                        continue
+                    else:
+                        raise ValueError(f"Configuration validation failed after {max_retries} attempts: {error_msg}")
+                
+                # Ensure we have exactly max_agents number of agents
+                if len(config.agents) > self.max_agents:
+                    config.agents = config.agents[:self.max_agents]
+                elif len(config.agents) < self.max_agents:
+                    logging.warning(f"Generated {len(config.agents)} agents, expected {self.max_agents}")
+                
+                return config
+                
+            except ValueError as e:
+                # Re-raise validation errors
+                raise
+            except Exception as e:
+                last_error = str(e)
+                if attempt < max_retries - 1:
+                    logging.warning(f"Error generating configuration (attempt {attempt + 1}/{max_retries}): {e}")
+                    continue
+                else:
+                    logging.error(f"Error generating configuration after {max_retries} attempts: {e}")
+                    raise
+        
+        # This should never be reached due to the raise statements above
+        raise RuntimeError(f"Failed to generate valid configuration after {max_retries} attempts")
 
     def _create_agents_and_tasks(self, config: AutoAgentsConfig) -> Tuple[List[Agent], List[Task]]:
         """Create agents and tasks from configuration"""
@@ -308,7 +460,9 @@ Return the configuration in a structured JSON format matching the AutoAgentsConf
                 max_rpm=self.max_rpm,
                 max_execution_time=self.max_execution_time,
                 max_iter=self.max_iter,
-                reflect_llm=self.reflect_llm
+                reflect_llm=self.reflect_llm,
+                base_url=self.base_url,
+                api_key=self.api_key
             )
             agents.append(agent)
             

{praisonaiagents-0.0.118 → praisonaiagents-0.0.123}/praisonaiagents/llm/__init__.py

@@ -32,6 +32,10 @@ from .openai_client import (
     ToolCall,
     process_stream_chunks
 )
+from .model_capabilities import (
+    supports_structured_outputs,
+    supports_streaming_with_tools
+)
 
 # Ensure telemetry is disabled after import as well
 try:
@@ -52,5 +56,7 @@ __all__ = [
     "CompletionUsage",
     "ChatCompletion",
     "ToolCall",
-    "process_stream_chunks"
+    "process_stream_chunks",
+    "supports_structured_outputs",
+    "supports_streaming_with_tools"
 ]