isa-model 0.3.91__py3-none-any.whl → 0.4.3__py3-none-any.whl

isa_model/inference/services/llm/openai_llm_service.py

@@ -1,6 +1,7 @@
 import logging
 import os
 import json
+import asyncio
 from typing import Dict, Any, List, Union, AsyncGenerator, Optional, Callable
 
 # 使用官方 OpenAI 库
@@ -17,9 +18,18 @@ class OpenAILLMService(BaseLLMService):
     def __init__(self, model_name: str = "gpt-4o-mini", provider_name: str = "openai", **kwargs):
         super().__init__(provider_name, model_name, **kwargs)
         
+        # Check if this is an O-series reasoning model
+        self.is_reasoning_model = model_name.startswith("o4-") or model_name.startswith("o3-")
+        self.uses_completion_tokens = self.is_reasoning_model or model_name.startswith("gpt-5")
+        self.requires_default_temperature = self.is_reasoning_model or model_name.startswith("gpt-5")
+        self.supports_deep_research = "deep-search" in model_name or "deep-research" in model_name
+        
         # Get configuration from centralized config manager
         provider_config = self.get_provider_config()
         
+        # Check if reasoning summary is enabled (requires verified organization)
+        self.enable_reasoning_summary = provider_config.get("enable_reasoning_summary", False)
+        
         # Initialize AsyncOpenAI client with provider configuration
         try:
             if not provider_config.get("api_key"):
@@ -28,7 +38,9 @@ class OpenAILLMService(BaseLLMService):
             self.client = AsyncOpenAI(
                 api_key=provider_config["api_key"],
                 base_url=provider_config.get("api_base_url", "https://api.openai.com/v1"),
-                organization=provider_config.get("organization")
+                organization=provider_config.get("organization"),
+                timeout=10.0,  # 10 second timeout for first token (much faster than 600s default)
+                max_retries=2  # Retry on timeout
             )
             
             logger.info(f"Initialized OpenAILLMService with model {self.model_name} and endpoint {self.client.base_url}")
@@ -40,11 +52,42 @@ class OpenAILLMService(BaseLLMService):
         self.last_token_usage = {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
         self.total_token_usage = {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0, "requests_count": 0}
         
+        # For O-series models, track reasoning tokens separately
+        if self.is_reasoning_model:
+            self.last_token_usage["reasoning_tokens"] = 0
+            self.total_token_usage["reasoning_tokens"] = 0
+        
     
     def _create_bound_copy(self) -> 'OpenAILLMService':
         """Create a copy of this service for tool binding"""
-        bound_service = OpenAILLMService(self.model_name, self.provider_name)
-        bound_service._bound_tools = self._bound_tools.copy()
+        # Create new instance but bypass full initialization
+        bound_service = object.__new__(OpenAILLMService)
+        
+        # Copy all essential attributes from original service
+        bound_service.model_name = self.model_name
+        bound_service.provider_name = self.provider_name
+        bound_service.client = self.client  # Reuse the same OpenAI client
+        bound_service.last_token_usage = self.last_token_usage.copy()
+        bound_service.total_token_usage = self.total_token_usage.copy()
+        bound_service._bound_tools = self._bound_tools.copy() if self._bound_tools else []
+        bound_service.adapter_manager = self.adapter_manager  # Reuse adapter manager
+        
+        # Copy OpenAI-specific attributes
+        bound_service.is_reasoning_model = self.is_reasoning_model
+        bound_service.uses_completion_tokens = self.uses_completion_tokens
+        bound_service.requires_default_temperature = self.requires_default_temperature
+        bound_service.supports_deep_research = self.supports_deep_research
+        
+        # Copy base class attributes
+        bound_service.streaming = self.streaming
+        bound_service.max_tokens = self.max_tokens
+        bound_service.temperature = self.temperature
+        bound_service._tool_mappings = {}
+        
+        # Copy BaseService attributes that are needed
+        bound_service.config_manager = self.config_manager
+        bound_service.model_manager = self.model_manager
+        
         return bound_service
     
     def bind_tools(self, tools: List[Any], **kwargs) -> 'OpenAILLMService':
@@ -66,16 +109,133 @@ class OpenAILLMService(BaseLLMService):
         
         return bound_service
     
-    async def astream(self, input_data: Union[str, List[Dict[str, str]], Any]) -> AsyncGenerator[str, None]:
+    async def astream(self, input_data: Union[str, List[Dict[str, str]], Any], show_reasoning: bool = False, **extra_kwargs) -> AsyncGenerator[Union[str, Dict[str, Any]], None]:
         """
         True streaming method - yields tokens one by one as they arrive
         
         Args:
             input_data: Same as ainvoke
+            show_reasoning: If True and model supports it, show reasoning process using Responses API
             
         Yields:
-            Individual tokens as they arrive from the API
+            Individual tokens as they arrive from the API, plus final result object with tool_calls
         """
+        try:
+            # Determine which API to use for streaming
+            use_responses_api = (show_reasoning and self.is_reasoning_model) or self.supports_deep_research
+            
+            if use_responses_api:
+                logger.info(f"Using Responses API streaming for {self.model_name}")
+                # Use Responses API streaming
+                async for chunk in self._astream_responses_api(input_data, show_reasoning, **extra_kwargs):
+                    yield chunk
+            else:
+                logger.debug(f"Using Chat Completions API streaming for {self.model_name}")
+                # Use Chat Completions API streaming
+                async for chunk in self._astream_chat_completions_api(input_data, **extra_kwargs):
+                    yield chunk
+                    
+        except Exception as e:
+            logger.error(f"Error in astream: {e}")
+            raise
+    
+    async def _astream_responses_api(self, input_data: Union[str, List[Dict[str, str]], Any], show_reasoning: bool = False, **extra_kwargs) -> AsyncGenerator[Union[str, Dict[str, Any]], None]:
+        """Stream using Responses API for reasoning models and deep research models"""
+        try:
+            # Use adapter manager to prepare messages
+            messages = self._prepare_messages(input_data)
+            
+            # Prepare request kwargs for Responses API
+            provider_config = self.get_provider_config()
+            kwargs = {
+                "model": self.model_name,
+                "input": messages,  # Responses API uses 'input' instead of 'messages'
+                "stream": True
+            }
+            
+            # Responses API uses max_output_tokens
+            max_tokens_value = provider_config.get("max_tokens", 1024)
+            kwargs["max_output_tokens"] = max_tokens_value
+            
+            # Add reasoning configuration if needed (optional - requires verified organization)
+            if show_reasoning and self.is_reasoning_model and self.enable_reasoning_summary:
+                kwargs["reasoning"] = {"summary": "auto"}
+                logger.info("Reasoning summary enabled - using verified organization features")
+            elif show_reasoning and self.is_reasoning_model:
+                logger.info("Reasoning visibility requested - using Responses API without summary (requires verified org)")
+            
+            # Deep research models require web_search_preview tool
+            if self.supports_deep_research:
+                kwargs["tools"] = [{"type": "web_search_preview"}]
+            
+            # Add any additional bound tools
+            tool_schemas = await self._prepare_tools_for_request()
+            if tool_schemas:
+                if "tools" not in kwargs:
+                    kwargs["tools"] = []
+                kwargs["tools"].extend(tool_schemas)
+            
+            # Stream using Responses API
+            content_chunks = []
+            reasoning_items = []
+            
+            try:
+                logger.info(f"Streaming with Responses API for model {self.model_name}")
+                stream = await self.client.responses.create(**kwargs)
+                
+                async for event in stream:
+                    # Handle different event types from Responses API
+                    if event.type == 'response.output_text.delta':
+                        # Stream text content
+                        if event.delta:
+                            content_chunks.append(event.delta)
+                            yield event.delta
+                    
+                    elif event.type == 'response.reasoning.delta' and show_reasoning:
+                        # Stream reasoning content (if enabled)
+                        if hasattr(event, 'delta') and event.delta:
+                            yield f"[思考: {event.delta}]"
+                    
+                    elif event.type == 'response.output_item.done':
+                        # Handle completed items (reasoning, function calls, etc.)
+                        if hasattr(event, 'item'):
+                            if event.item.type == 'reasoning':
+                                reasoning_items.append(event.item)
+                            elif event.item.type == 'function_call':
+                                # Handle function call completion
+                                logger.debug(f"Function call completed: {event.item}")
+                
+                # Create final response object
+                full_content = "".join(content_chunks)
+                
+                # Track usage for streaming
+                self._track_streaming_usage(messages, full_content)
+                
+                # Get billing info
+                await asyncio.sleep(0.01)
+                billing_info = self._get_streaming_billing_info()
+                
+                # Format final result
+                final_result = self._format_response(full_content, input_data)
+                
+                # Yield final result with metadata
+                yield {
+                    "result": final_result,
+                    "billing": billing_info,
+                    "reasoning_items": len(reasoning_items),
+                    "api_used": "responses"
+                }
+                
+            except Exception as e:
+                logger.error(f"Error in Responses API streaming: {e}")
+                raise
+                
+        except Exception as e:
+            logger.error(f"Error in _astream_responses_api: {e}")
+            raise
+    
+    async def _astream_chat_completions_api(self, input_data: Union[str, List[Dict[str, str]], Any], **extra_kwargs) -> AsyncGenerator[Union[str, Dict[str, Any]], None]:
+        """Stream using Chat Completions API for standard models"""
         try:
             # Use adapter manager to prepare messages
             messages = self._prepare_messages(input_data)
@@ -85,86 +245,275 @@ class OpenAILLMService(BaseLLMService):
             kwargs = {
                 "model": self.model_name,
                 "messages": messages,
-                "temperature": provider_config.get("temperature", 0.7),
-                "max_tokens": provider_config.get("max_tokens", 1024),
                 "stream": True
             }
             
+            # O4 and GPT-5 models only support temperature=1 (default)
+            if not self.requires_default_temperature:
+                kwargs["temperature"] = provider_config.get("temperature", 0.7)
+            
+            # O4 and GPT-5 models use max_completion_tokens instead of max_tokens
+            max_tokens_value = provider_config.get("max_tokens", 1024)
+            if self.uses_completion_tokens:
+                kwargs["max_completion_tokens"] = max_tokens_value
+            else:
+                kwargs["max_tokens"] = max_tokens_value
+            
             # Add tools if bound using adapter manager
             tool_schemas = await self._prepare_tools_for_request()
             if tool_schemas:
                 kwargs["tools"] = tool_schemas
                 kwargs["tool_choice"] = "auto"
             
-            # Stream tokens one by one
+            # Add response_format if specified (for JSON mode)
+            if 'response_format' in extra_kwargs:
+                kwargs['response_format'] = extra_kwargs['response_format']
+                logger.debug(f"Using response_format in streaming: {extra_kwargs['response_format']}")
+            
+            # Stream tokens and detect tool calls
             content_chunks = []
+            tool_calls_accumulator = {}  # Track complete tool calls by ID
+            has_tool_calls = False
+            
             try:
                 stream = await self.client.chat.completions.create(**kwargs)
                 async for chunk in stream:
-                    content = chunk.choices[0].delta.content
-                    if content:
-                        content_chunks.append(content)
-                        yield content
+                    delta = chunk.choices[0].delta
+                    
+                    # Check for tool calls first
+                    if hasattr(delta, 'tool_calls') and delta.tool_calls:
+                        has_tool_calls = True
+                        for tool_call in delta.tool_calls:
+                            tool_index = getattr(tool_call, 'index', 0)  # OpenAI uses index for streaming
+                            
+                            # Use index as key since streaming tool calls use index
+                            tool_key = f"tool_{tool_index}"
+                            
+                            # Initialize tool call if not seen before
+                            if tool_key not in tool_calls_accumulator:
+                                tool_calls_accumulator[tool_key] = {
+                                    'id': getattr(tool_call, 'id', f"call_{tool_index}"),
+                                    'type': 'function',
+                                    'function': {
+                                        'name': '',
+                                        'arguments': ''
+                                    }
+                                }
+                            
+                            # Accumulate function name
+                            if hasattr(tool_call, 'function') and hasattr(tool_call.function, 'name') and tool_call.function.name:
+                                tool_calls_accumulator[tool_key]['function']['name'] += tool_call.function.name
+                            
+                            # Accumulate function arguments
+                            if hasattr(tool_call, 'function') and hasattr(tool_call.function, 'arguments'):
+                                if tool_call.function.arguments:
+                                    tool_calls_accumulator[tool_key]['function']['arguments'] += tool_call.function.arguments
+                    
+                    # Handle regular content - only stream if no tool calls detected
+                    elif delta.content:
+                        content_chunks.append(delta.content)
+                        if not has_tool_calls:  # Only yield content if no tool calls
+                            yield delta.content
+                
+                # Always yield final result at the end
+                # - If has tool_calls: complete structured response (no prior streaming)
+                # - If no tool_calls: AIMessage after streaming content
+                
+                # Create a mock message object for adapter processing
+                class MockMessage:
+                    def __init__(self):
+                        self.content = "".join(content_chunks) or ""
+                        self.tool_calls = []
+                        # Add tool_calls if any
+                        if tool_calls_accumulator:
+                            for tool_data in tool_calls_accumulator.values():
+                                mock_tool_call = type('MockToolCall', (), {
+                                    'id': tool_data['id'],
+                                    'function': type('MockFunction', (), {
+                                        'name': tool_data['function']['name'],
+                                        'arguments': tool_data['function']['arguments']
+                                    })()
+                                })()
+                                self.tool_calls.append(mock_tool_call)
+                
+                mock_message = MockMessage()
+                
+                logger.debug(f"Streaming complete - tool calls collected: {len(mock_message.tool_calls)}")
+                for i, tc in enumerate(mock_message.tool_calls):
+                    logger.debug(f"  Tool call {i+1}: {tc.function.name} with args: {tc.function.arguments}")
+                
+                # Format response using adapter (this handles LangChain conversion)
+                final_result = self._format_response(mock_message, input_data)
+                
+                logger.debug(f"Final result type after adapter: {type(final_result)}")
+                logger.debug(f"Final result has tool_calls: {hasattr(final_result, 'tool_calls')}")
                 
                 # Track usage after streaming is complete
                 full_content = "".join(content_chunks)
                 self._track_streaming_usage(messages, full_content)
                 
+                # Get billing info after tracking (wait a moment for billing to be recorded)
+                await asyncio.sleep(0.01)
+                billing_info = self._get_streaming_billing_info()
+                
+                # Yield the final result with billing info
+                yield {
+                    "result": final_result,
+                    "billing": billing_info,
+                    "api_used": "chat_completions"
+                }
+                
             except Exception as e:
-                logger.error(f"Error in streaming: {e}")
+                logger.error(f"Error in Chat Completions streaming: {e}")
                 raise
             
         except Exception as e:
-            logger.error(f"Error in astream: {e}")
+            logger.error(f"Error in _astream_chat_completions_api: {e}")
             raise
     
-    async def ainvoke(self, input_data: Union[str, List[Dict[str, str]], Any]) -> Union[str, Any]:
-        """Unified invoke method for all input types"""
+    async def ainvoke(self, input_data: Union[str, List[Dict[str, str]], Any], show_reasoning: bool = False, **extra_kwargs) -> Union[str, Any]:
+        """
+        Unified invoke method for all input types
+        
+        Args:
+            input_data: Input messages or text
+            show_reasoning: If True and model supports it, show reasoning process using Responses API
+            **extra_kwargs: Additional parameters to pass to the API (e.g., response_format)
+        """
         try:
             # Use adapter manager to prepare messages
             messages = self._prepare_messages(input_data)
             
+            # Determine which API to use
+            # Responses API is required for:
+            # 1. Reasoning models with show_reasoning=True
+            # 2. Deep research models (they only work with Responses API)
+            use_responses_api = (show_reasoning and self.is_reasoning_model) or self.supports_deep_research
+            
             # Prepare request kwargs
             provider_config = self.get_provider_config()
             kwargs = {
                 "model": self.model_name,
-                "messages": messages,
-                "temperature": provider_config.get("temperature", 0.7),
-                "max_tokens": provider_config.get("max_tokens", 1024)
+                "messages": messages
             }
             
+            # O4 and GPT-5 models only support temperature=1 (default)
+            if not self.requires_default_temperature:
+                kwargs["temperature"] = provider_config.get("temperature", 0.7)
+            
+            # O4 and GPT-5 models use max_completion_tokens instead of max_tokens
+            max_tokens_value = provider_config.get("max_tokens", 1024)
+            if self.uses_completion_tokens:
+                kwargs["max_completion_tokens"] = max_tokens_value
+            else:
+                kwargs["max_tokens"] = max_tokens_value
+            
             # Add tools if bound using adapter manager
             tool_schemas = await self._prepare_tools_for_request()
             if tool_schemas:
                 kwargs["tools"] = tool_schemas
-                kwargs["tool_choice"] = "auto"
+                if not use_responses_api:  # Responses API handles tool choice differently
+                    kwargs["tool_choice"] = "auto"
+            
+            # Add response_format if specified (for JSON mode)
+            if 'response_format' in extra_kwargs:
+                kwargs['response_format'] = extra_kwargs['response_format']
+                logger.debug(f"Using response_format: {extra_kwargs['response_format']}")
             
             # Handle streaming vs non-streaming
             if self.streaming:
                 # TRUE STREAMING MODE - collect all chunks from the stream
                 content_chunks = []
-                async for token in self.astream(input_data):
-                    content_chunks.append(token)
-                content = "".join(content_chunks)
+                async for token in self.astream(input_data, show_reasoning=show_reasoning, **extra_kwargs):
+                    if isinstance(token, str):
+                        content_chunks.append(token)
+                    elif isinstance(token, dict) and "result" in token:
+                        # Return the final result from streaming
+                        return token["result"]
                 
+                # Fallback: join collected content
+                content = "".join(content_chunks)
                 return self._format_response(content, input_data)
             else:
-                # Non-streaming mode
-                response = await self.client.chat.completions.create(**kwargs)
-                message = response.choices[0].message
-                
-                # Update usage tracking
-                if response.usage:
-                    self._update_token_usage(response.usage)
-                    await self._track_billing(response.usage)
-                
-                # Handle tool calls if present - let adapter process the complete message
-                if message.tool_calls:
-                    # Pass the complete message object to adapter for proper tool_calls handling
-                    return self._format_response(message, input_data)
-                
-                # Return appropriate format based on input type
-                return self._format_response(message.content or "", input_data)
+                # Non-streaming mode - choose API based on reasoning visibility
+                if use_responses_api:
+                    logger.info(f"Using Responses API for model {self.model_name}")
+                    
+                    # Convert kwargs for Responses API
+                    responses_kwargs = {
+                        "model": kwargs["model"],
+                        "input": kwargs["messages"]  # Responses API uses 'input' instead of 'messages'
+                    }
+                    
+                    # Handle max tokens parameter
+                    if "max_completion_tokens" in kwargs:
+                        responses_kwargs["max_output_tokens"] = kwargs["max_completion_tokens"]
+                    elif "max_tokens" in kwargs:
+                        responses_kwargs["max_output_tokens"] = kwargs["max_tokens"]
+                    
+                    # Add tools if present
+                    if "tools" in kwargs:
+                        responses_kwargs["tools"] = kwargs["tools"]
+                    
+                    # Add reasoning configuration for reasoning models (requires verified organization)
+                    if show_reasoning and self.is_reasoning_model and self.enable_reasoning_summary:
+                        responses_kwargs["reasoning"] = {"summary": "auto"}
+                        logger.info("Reasoning summary enabled - using verified organization features")
+                    elif show_reasoning and self.is_reasoning_model:
+                        logger.info("Reasoning visibility requested - using Responses API without summary (requires verified org)")
+                    
+                    # Deep research models require web_search_preview tool
+                    if self.supports_deep_research:
+                        if "tools" not in responses_kwargs:
+                            responses_kwargs["tools"] = []
+                        responses_kwargs["tools"].insert(0, {"type": "web_search_preview"})
+                    
+                    response = await self.client.responses.create(**responses_kwargs)
+                    
+                    # Handle Responses API format
+                    if hasattr(response, 'output_text'):
+                        # Modern Responses API format
+                        content = response.output_text
+                        usage_info = getattr(response, 'usage', None)
+                    elif hasattr(response, 'body') and hasattr(response.body, 'response'):
+                        # Legacy format
+                        content = response.body.response
+                        usage_info = getattr(response.body, 'usage', None)
+                    else:
+                        # Fallback handling
+                        content = str(response)
+                        usage_info = None
+                    
+                    # Update usage tracking if available
+                    if usage_info:
+                        self._update_token_usage(usage_info)
+                        await self._track_billing(usage_info)
+                    
+                    return self._format_response(content, input_data)
+                else:
+                    # Standard Chat Completions API
+                    response = await self.client.chat.completions.create(**kwargs)
+                    message = response.choices[0].message
+                    
+                    # Debug: Log the raw OpenAI response
+                    logger.debug(f"OpenAI response message: {message}")
+                    if message.tool_calls:
+                        logger.debug(f"Tool calls found: {len(message.tool_calls)}")
+                        for i, tc in enumerate(message.tool_calls):
+                            logger.debug(f"  Tool call {i+1}: id={tc.id}, function={tc.function.name}, args={tc.function.arguments}")
+                    
+                    # Update usage tracking
+                    if response.usage:
+                        self._update_token_usage(response.usage)
+                        await self._track_billing(response.usage)
+                    
+                    # Handle tool calls if present - let adapter process the complete message
+                    if message.tool_calls:
+                        # Pass the complete message object to adapter for proper tool_calls handling
+                        return self._format_response(message, input_data)
+                    
+                    # Return appropriate format based on input type
+                    return self._format_response(message.content or "", input_data)
             
         except Exception as e:
             logger.error(f"Error in ainvoke: {e}")
@@ -210,11 +559,42 @@ class OpenAILLMService(BaseLLMService):
     
     def _update_token_usage(self, usage):
         """Update token usage statistics"""
-        self.last_token_usage = {
-            "prompt_tokens": usage.prompt_tokens,
-            "completion_tokens": usage.completion_tokens,
-            "total_tokens": usage.total_tokens
-        }
+        # Handle different usage object structures (Chat Completions vs Responses API)
+        if hasattr(usage, 'prompt_tokens'):
+            # Chat Completions API format
+            self.last_token_usage = {
+                "prompt_tokens": usage.prompt_tokens,
+                "completion_tokens": usage.completion_tokens,
+                "total_tokens": usage.total_tokens
+            }
+        elif hasattr(usage, 'input_tokens'):
+            # Responses API format
+            self.last_token_usage = {
+                "prompt_tokens": usage.input_tokens,
+                "completion_tokens": usage.output_tokens,
+                "total_tokens": usage.total_tokens
+            }
+        else:
+            # Fallback for unknown usage format
+            logger.warning(f"Unknown usage format: {type(usage)}, attributes: {dir(usage)}")
+            self.last_token_usage = {
+                "prompt_tokens": 0,
+                "completion_tokens": 0,
+                "total_tokens": 0
+            }
+        
+        # For O-series models, track reasoning tokens if available
+        if self.is_reasoning_model:
+            reasoning_tokens = 0
+            if hasattr(usage, 'reasoning_tokens'):
+                reasoning_tokens = usage.reasoning_tokens
+            elif hasattr(usage, 'output_tokens_details') and hasattr(usage.output_tokens_details, 'reasoning_tokens'):
+                reasoning_tokens = usage.output_tokens_details.reasoning_tokens
+            
+            self.last_token_usage["reasoning_tokens"] = reasoning_tokens
+            if "reasoning_tokens" not in self.total_token_usage:
+                self.total_token_usage["reasoning_tokens"] = 0
+            self.total_token_usage["reasoning_tokens"] += reasoning_tokens
         
         # Update total usage
         self.total_token_usage["prompt_tokens"] += self.last_token_usage["prompt_tokens"]
@@ -225,15 +605,35 @@ class OpenAILLMService(BaseLLMService):
     async def _track_billing(self, usage):
         """Track billing information"""
         provider_config = self.get_provider_config()
+        
+        # Prepare metadata for tracking
+        metadata = {
+            "temperature": provider_config.get("temperature", 0.7), 
+            "max_tokens": provider_config.get("max_tokens", 1024),
+            "is_reasoning_model": self.is_reasoning_model
+        }
+        
+        # Add reasoning tokens if available for O-series models
+        if self.is_reasoning_model and hasattr(usage, 'reasoning_tokens'):
+            metadata["reasoning_tokens"] = usage.reasoning_tokens
+        
+        # Get tokens using the same logic as _update_token_usage
+        if hasattr(usage, 'prompt_tokens'):
+            input_tokens = usage.prompt_tokens
+            output_tokens = usage.completion_tokens
+        elif hasattr(usage, 'input_tokens'):
+            input_tokens = usage.input_tokens
+            output_tokens = usage.output_tokens
+        else:
+            input_tokens = 0
+            output_tokens = 0
+            
         await self._track_usage(
             service_type=ServiceType.LLM,
             operation="chat",
-            input_tokens=usage.prompt_tokens,
-            output_tokens=usage.completion_tokens,
-            metadata={
-                "temperature": provider_config.get("temperature", 0.7), 
-                "max_tokens": provider_config.get("max_tokens", 1024)
-            }
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+            metadata=metadata
         )
     
     def get_token_usage(self) -> Dict[str, Any]:
@@ -252,14 +652,18 @@ class OpenAILLMService(BaseLLMService):
             "max_tokens": provider_config.get("max_tokens", 1024),
             "supports_streaming": True,
             "supports_functions": True,
-            "provider": "openai"
+            "supports_reasoning": self.is_reasoning_model,
+            "supports_deep_research": self.supports_deep_research,
+            "provider": "openai",
+            "model_type": "reasoning" if self.is_reasoning_model else "standard"
         }
     
         
     async def chat(
         self,
         input_data: Union[str, List[Dict[str, str]], Any],
-        max_tokens: Optional[int] = None
+        max_tokens: Optional[int] = None,
+        show_reasoning: bool = False
     ) -> Dict[str, Any]:
         """
         Chat method that wraps ainvoke for compatibility with base class
@@ -267,13 +671,14 @@ class OpenAILLMService(BaseLLMService):
         Args:
             input_data: Input messages
             max_tokens: Maximum tokens to generate
+            show_reasoning: Whether to show reasoning process (for O4 models)
             
         Returns:
             Dict containing chat response with properly formatted message object
         """
         try:
-            # Call ainvoke and get the response (already processed by adapter)
-            response = await self.ainvoke(input_data)
+            # Call ainvoke with show_reasoning parameter
+            response = await self.ainvoke(input_data, show_reasoning=show_reasoning)
             
             # Return the response as-is (adapter already formatted it correctly)
             # For LangChain inputs, this will be an AIMessage object
@@ -284,7 +689,9 @@ class OpenAILLMService(BaseLLMService):
                 "metadata": {
                     "model": self.model_name,
                     "provider": self.provider_name,
-                    "max_tokens": max_tokens or self.max_tokens
+                    "max_tokens": max_tokens or self.max_tokens,
+                    "show_reasoning": show_reasoning,
+                    "is_reasoning_model": self.is_reasoning_model
                 }
             }
         except Exception as e:
@@ -299,6 +706,213 @@ class OpenAILLMService(BaseLLMService):
                 }
             }
     
+    async def deep_research(
+        self,
+        input_data: Union[str, List[Dict[str, str]], Any],
+        research_type: Optional[str] = None,
+        search_enabled: bool = True
+    ) -> Dict[str, Any]:
+        """
+        深度研究任务 - 专为深度研究模型设计，使用OpenAI Responses API
+        
+        Args:
+            input_data: 研究查询或问题
+            research_type: 研究类型 (academic, market, competitive, etc.)
+            search_enabled: 是否启用网络搜索
+            
+        Returns:
+            Dict containing research results
+        """
+        if not self.supports_deep_research:
+            # Fallback to regular chat for non-deep-research models
+            logger.info(f"Model {self.model_name} doesn't support deep research, falling back to regular chat")
+            return await self.chat(input_data)
+        
+        try:
+            # Prepare messages with research context
+            messages = self._prepare_messages(input_data)
+            
+            # Add research-specific system prompt if research_type is specified
+            if research_type and messages:
+                research_prompts = {
+                    "academic": "You are conducting academic research. Please provide thorough, well-sourced analysis with proper citations and methodical reasoning.",
+                    "market": "You are conducting market research. Focus on market trends, competitive analysis, and business insights.",
+                    "competitive": "You are conducting competitive analysis. Compare and contrast different approaches, solutions, or entities.",
+                    "technical": "You are conducting technical research. Provide detailed technical analysis with implementation considerations."
+                }
+                
+                if research_type in research_prompts:
+                    # Insert system message at the beginning
+                    system_msg = {"role": "system", "content": research_prompts[research_type]}
+                    if messages[0].get("role") == "system":
+                        messages[0]["content"] = research_prompts[research_type] + "\n\n" + messages[0]["content"]
+                    else:
+                        messages.insert(0, system_msg)
+            
+            # Prepare request kwargs for Responses API
+            provider_config = self.get_provider_config()
+            kwargs = {
+                "model": self.model_name,
+                "input": messages  # Responses API uses 'input' instead of 'messages'
+            }
+            
+            # Responses API uses max_output_tokens instead of max_completion_tokens
+            max_tokens_value = provider_config.get("max_tokens", 4096)
+            kwargs["max_output_tokens"] = max_tokens_value
+            
+            # Deep research models require web_search_preview tool when search is enabled
+            if search_enabled:
+                kwargs["tools"] = [
+                    {
+                        "type": "web_search_preview"
+                    }
+                ]
+            
+            # Add any additional bound tools
+            tool_schemas = await self._prepare_tools_for_request()
+            if tool_schemas:
+                if "tools" not in kwargs:
+                    kwargs["tools"] = []
+                kwargs["tools"].extend(tool_schemas)
+            
+            # Check if streaming is enabled
+            if self.streaming:
+                # Use streaming mode for deep research
+                logger.info(f"Using Responses API streaming for deep research model {self.model_name}")
+                kwargs["stream"] = True
+                
+                content_chunks = []
+                stream = await self.client.responses.create(**kwargs)
+                
+                async for event in stream:
+                    if event.type == 'response.output_text.delta':
+                        if event.delta:
+                            content_chunks.append(event.delta)
+                
+                message_content = "".join(content_chunks)
+                
+                # Track estimated usage for streaming
+                messages = self._prepare_messages(input_data)
+                self._track_streaming_usage(messages, message_content)
+                
+                # Format response
+                formatted_response = self._format_response(message_content or "", input_data)
+            else:
+                # Use non-streaming mode for deep research
+                logger.info(f"Using Responses API for deep research model {self.model_name}")
+                response = await self.client.responses.create(**kwargs)
+                
+                # Extract the response content from Responses API format
+                if hasattr(response, 'output_text'):
+                    # Modern Responses API format
+                    message_content = response.output_text
+                    usage_info = getattr(response, 'usage', None)
+                elif hasattr(response, 'body') and hasattr(response.body, 'response'):
+                    # Legacy Responses API format
+                    message_content = response.body.response
+                    usage_info = getattr(response.body, 'usage', None)
+                elif hasattr(response, 'choices') and response.choices:
+                    # Fallback to standard format
+                    message_content = response.choices[0].message.content
+                    usage_info = getattr(response, 'usage', None)
+                else:
+                    # Handle unexpected format
+                    message_content = str(response)
+                    usage_info = None
+                
+                # Update usage tracking if available
+                if usage_info:
+                    self._update_token_usage(usage_info)
+                    await self._track_billing(usage_info)
+                
+                # Format response
+                formatted_response = self._format_response(message_content or "", input_data)
+            
+            return {
+                "result": formatted_response,
+                "research_type": research_type,
+                "search_enabled": search_enabled,
+                "success": True,
+                "metadata": {
+                    "model": self.model_name,
+                    "provider": self.provider_name,
+                    "supports_deep_research": self.supports_deep_research,
+                    "reasoning_model": self.is_reasoning_model,
+                    "api_used": "responses"
+                }
+            }
+            
+        except Exception as e:
+            logger.error(f"Deep research failed: {e}")
+            return {
+                "result": None,
+                "success": False,
+                "error": str(e),
+                "metadata": {
+                    "model": self.model_name,
+                    "provider": self.provider_name,
+                    "api_used": "responses"
+                }
+            }
+
     async def close(self):
         """Close the backend client"""
-        await self.client.close()
+        await self.client.close()
+    
+    def _get_streaming_billing_info(self) -> Dict[str, Any]:
+        """Get billing information for streaming requests"""
+        try:
+            # Check if service has model_manager with billing_tracker
+            if hasattr(self, 'model_manager') and hasattr(self.model_manager, 'billing_tracker'):
+                billing_tracker = self.model_manager.billing_tracker
+                
+                # Get the latest usage record for this model
+                model_records = [
+                    record for record in billing_tracker.usage_records
+                    if record.model_id == self.model_name
+                ]
+                
+                if model_records:
+                    # Get the most recent record
+                    latest_record = max(model_records, key=lambda r: r.timestamp)
+                    
+                    return {
+                        "cost_usd": latest_record.cost_usd,
+                        "input_tokens": latest_record.input_tokens,
+                        "output_tokens": latest_record.output_tokens,
+                        "total_tokens": latest_record.total_tokens,
+                        "operation": latest_record.operation,
+                        "timestamp": latest_record.timestamp,
+                        "currency": "USD"
+                    }
+            
+            # Fallback: use last token usage with estimated cost
+            last_usage = self.get_last_token_usage()
+            estimated_cost = 0.0
+            
+            if hasattr(self, 'model_manager'):
+                estimated_cost = self.model_manager.calculate_cost(
+                    provider=self.provider_name,
+                    model_name=self.model_name,
+                    input_tokens=last_usage.get("prompt_tokens", 0),
+                    output_tokens=last_usage.get("completion_tokens", 0)
+                )
+            
+            return {
+                "cost_usd": estimated_cost,
+                "input_tokens": last_usage.get("prompt_tokens", 0),
+                "output_tokens": last_usage.get("completion_tokens", 0),
+                "total_tokens": last_usage.get("total_tokens", 0),
+                "operation": "chat",
+                "timestamp": None,
+                "currency": "USD",
+                "note": "Estimated from last token usage"
+            }
+            
+        except Exception as e:
+            logger.warning(f"Failed to get streaming billing info: {e}")
+            return {
+                "cost_usd": 0.0,
+                "error": str(e),
+                "currency": "USD"
+            }