evalscope 0.17.1__py3-none-any.whl → 1.0.0__py3-none-any.whl

evalscope/models/utils/openai.py

@@ -0,0 +1,698 @@
+import base64
+import json
+import re
+from collections import defaultdict
+from copy import copy
+from openai import APIStatusError, OpenAIError
+from openai.types.chat import (
+    ChatCompletion,
+    ChatCompletionAssistantMessageParam,
+    ChatCompletionChunk,
+    ChatCompletionContentPartImageParam,
+    ChatCompletionContentPartInputAudioParam,
+    ChatCompletionContentPartParam,
+    ChatCompletionContentPartRefusalParam,
+    ChatCompletionContentPartTextParam,
+    ChatCompletionDeveloperMessageParam,
+    ChatCompletionMessage,
+    ChatCompletionMessageParam,
+    ChatCompletionMessageToolCall,
+    ChatCompletionMessageToolCallParam,
+    ChatCompletionNamedToolChoiceParam,
+    ChatCompletionSystemMessageParam,
+    ChatCompletionToolChoiceOptionParam,
+    ChatCompletionToolMessageParam,
+    ChatCompletionToolParam,
+    ChatCompletionUserMessageParam,
+)
+from openai.types.chat.chat_completion import Choice, ChoiceLogprobs
+from openai.types.chat.chat_completion_message_tool_call import Function
+from openai.types.completion_usage import CompletionUsage
+from openai.types.shared_params.function_definition import FunctionDefinition
+from pydantic import JsonValue
+from typing import Any, Dict, List, Literal, Optional, Tuple, Union
+
+from evalscope.api.messages import (
+    ChatMessage,
+    ChatMessageAssistant,
+    ChatMessageSystem,
+    ChatMessageTool,
+    ChatMessageUser,
+    Content,
+    ContentAudio,
+    ContentImage,
+    ContentReasoning,
+    ContentText,
+    parse_content_with_reasoning,
+)
+from evalscope.api.model import (
+    ChatCompletionChoice,
+    GenerateConfig,
+    Logprobs,
+    ModelOutput,
+    ModelUsage,
+    StopReason,
+    as_stop_reason,
+)
+from evalscope.api.tool import ToolCall, ToolChoice, ToolFunction, ToolInfo, parse_tool_call
+from evalscope.utils.url_utils import file_as_data_uri, is_http_url
+
+BASE_64_DATA_REMOVED = '<base64-data-removed>'
+
+
+class OpenAIResponseError(OpenAIError):
+
+    def __init__(self, code: str, message: str) -> None:
+        self.code = code
+        self.message = message
+
+    def __str__(self) -> str:
+        return f'{self.code}: {self.message}'
+
+
+def openai_chat_tool_call(tool_call: ToolCall) -> ChatCompletionMessageToolCall:
+    return ChatCompletionMessageToolCall(
+        type='function',
+        id=tool_call.id,
+        function=Function(name=tool_call.function.name, arguments=json.dumps(tool_call.function.arguments)),
+    )
+
+
+def openai_chat_tool_call_param(tool_call: ToolCall) -> ChatCompletionMessageToolCallParam:
+    return ChatCompletionMessageToolCallParam(
+        id=tool_call.id,
+        function=dict(name=tool_call.function.name, arguments=json.dumps(tool_call.function.arguments)),
+        type='function',
+    )
+
+
+def openai_chat_completion_part(content: Content) -> ChatCompletionContentPartParam:
+    if content.type == 'text':
+        return ChatCompletionContentPartTextParam(type='text', text=content.text)
+    elif content.type == 'image':
+        # API takes URL or base64 encoded file. If it's a remote file or
+        # data URL leave it alone, otherwise encode it
+        image_url = content.image
+        detail = content.detail
+
+        if not is_http_url(image_url):
+            image_url = file_as_data_uri(image_url)
+
+        return ChatCompletionContentPartImageParam(
+            type='image_url',
+            image_url=dict(url=image_url, detail=detail),
+        )
+    elif content.type == 'audio':
+        audio_data_uri = file_as_data_uri(content.audio)
+        audio_data = audio_data_uri.split('base64,')[1]
+
+        return ChatCompletionContentPartInputAudioParam(
+            type='input_audio', input_audio=dict(data=audio_data, format=content.format)
+        )
+
+    else:
+        raise RuntimeError('Video content is not currently supported by Open AI chat models.')
+
+
+def openai_chat_message(
+    message: ChatMessage, system_role: Literal['user', 'system', 'developer'] = 'system'
+) -> ChatCompletionMessageParam:
+    if message.role == 'system':
+        if system_role == 'user':
+            return ChatCompletionUserMessageParam(role='user', content=message.text)
+        elif system_role == 'system':
+            return ChatCompletionSystemMessageParam(role=message.role, content=message.text)
+        elif system_role == 'developer':
+            return ChatCompletionDeveloperMessageParam(role='developer', content=message.text)
+    elif message.role == 'user':
+        return ChatCompletionUserMessageParam(
+            role=message.role,
+            content=(
+                message.content if isinstance(message.content, str) else
+                [openai_chat_completion_part(content) for content in message.content]
+            ),
+        )
+    elif message.role == 'assistant':
+        if message.tool_calls:
+            return ChatCompletionAssistantMessageParam(
+                role=message.role,
+                content=openai_assistant_content(message),
+                tool_calls=[openai_chat_tool_call_param(call) for call in message.tool_calls],
+            )
+        else:
+            return ChatCompletionAssistantMessageParam(role=message.role, content=openai_assistant_content(message))
+    elif message.role == 'tool':
+        return ChatCompletionToolMessageParam(
+            role=message.role,
+            content=(f'Error: {message.error.message}' if message.error else message.text),
+            tool_call_id=str(message.tool_call_id),
+        )
+    else:
+        raise ValueError(f'Unexpected message role {message.role}')
+
+
+def openai_chat_messages(
+    messages: List[ChatMessage],
+    system_role: Literal['user', 'system', 'developer'] = 'system',
+) -> List[ChatCompletionMessageParam]:
+    return [openai_chat_message(message, system_role) for message in messages]
+
+
+def openai_completion_params(model: str, config: GenerateConfig, tools: bool) -> Dict[str, Any]:
+    params: Dict[str, Any] = dict(model=model)
+    # handle stream option
+    if config.stream is not None:
+        params['stream'] = config.stream
+        if config.stream:
+            params['stream_options'] = {'include_usage': True}
+    if config.timeout is not None:
+        params['timeout'] = config.timeout
+    if config.max_tokens is not None:
+        params['max_tokens'] = config.max_tokens
+    if config.frequency_penalty is not None:
+        params['frequency_penalty'] = config.frequency_penalty
+    if config.stop_seqs is not None:
+        params['stop'] = config.stop_seqs
+    if config.presence_penalty is not None:
+        params['presence_penalty'] = config.presence_penalty
+    if config.logit_bias is not None:
+        params['logit_bias'] = config.logit_bias
+    if config.seed is not None:
+        params['seed'] = config.seed
+    if config.temperature is not None:
+        params['temperature'] = config.temperature
+    if config.top_p is not None:
+        params['top_p'] = config.top_p
+    if config.n is not None:
+        params['n'] = config.n
+    if config.logprobs is not None:
+        params['logprobs'] = config.logprobs
+    if config.top_logprobs is not None:
+        params['top_logprobs'] = config.top_logprobs
+    if tools and config.parallel_tool_calls is not None:
+        params['parallel_tool_calls'] = config.parallel_tool_calls
+    if config.reasoning_effort is not None:
+        params['reasoning_effort'] = config.reasoning_effort
+    if config.response_schema is not None:
+        params['response_format'] = dict(
+            type='json_schema',
+            json_schema=dict(
+                name=config.response_schema.name,
+                schema=config.response_schema.json_schema.model_dump(exclude_none=True),
+                description=config.response_schema.description,
+                strict=config.response_schema.strict,
+            ),
+        )
+    if config.extra_body:
+        params['extra_body'] = config.extra_body
+
+    return params
+
+
+def openai_assistant_content(message: ChatMessageAssistant) -> str:
+    # In agent bridge scenarios, we could encounter concepts such as reasoning and
+    # .internal use in the ChatMessageAssistant that are not supported by the OpenAI
+    # choices API. This code smuggles that data into the plain text so that it
+    # survives multi-turn round trips.
+
+    if isinstance(message.content, str):
+        content = message.content
+    else:
+        content = ''
+        for c in message.content:
+            if c.type == 'reasoning':
+                attribs = ''
+                if c.signature is not None:
+                    attribs = f'{attribs} signature="{c.signature}"'
+                if c.redacted:
+                    attribs = f'{attribs} redacted="true"'
+                content = f'{content}\n<think{attribs}>\n{c.reasoning}\n</think>\n'
+            elif c.type == 'text':
+                content = f'{content}\n{c.text}'
+
+    if message.internal:
+        content = f"""{content}\n<internal>{
+            base64.b64encode(json.dumps(message.internal).encode("utf-8")).decode(
+                "utf-8"
+            )
+        }</internal>\n"""
+    return content
+
+
+def openai_chat_choices(choices: List[ChatCompletionChoice]) -> List[Choice]:
+    oai_choices: List[Choice] = []
+
+    for index, choice in enumerate(choices):
+        content = openai_assistant_content(choice.message)
+        if choice.message.tool_calls:
+            tool_calls = [openai_chat_tool_call(tc) for tc in choice.message.tool_calls]
+        else:
+            tool_calls = None
+        message = ChatCompletionMessage(role='assistant', content=content, tool_calls=tool_calls)
+        oai_choices.append(
+            Choice(
+                finish_reason=openai_finish_reason(choice.stop_reason),
+                index=index,
+                message=message,
+                logprobs=ChoiceLogprobs(**choice.logprobs.model_dump()) if choice.logprobs is not None else None,
+            )
+        )
+
+    return oai_choices
+
+
+def openai_completion_usage(usage: ModelUsage) -> CompletionUsage:
+    return CompletionUsage(
+        completion_tokens=usage.output_tokens,
+        prompt_tokens=usage.input_tokens,
+        total_tokens=usage.total_tokens,
+    )
+
+
+def openai_finish_reason(
+    stop_reason: StopReason
+) -> Literal['stop', 'length', 'tool_calls', 'content_filter', 'function_call']:
+    if stop_reason in ('stop', 'tool_calls', 'content_filter'):
+        return stop_reason
+    elif stop_reason == 'model_length':
+        return 'length'
+    else:
+        return 'stop'
+
+
+def openai_chat_tool_param(tool: ToolInfo) -> ChatCompletionToolParam:
+    function = FunctionDefinition(
+        name=tool.name,
+        description=tool.description,
+        parameters=tool.parameters.model_dump(exclude_none=True),
+    )
+    return ChatCompletionToolParam(type='function', function=function)
+
+
+def openai_chat_tools(tools: List[ToolInfo]) -> List[ChatCompletionToolParam]:
+    return [openai_chat_tool_param(tool) for tool in tools]
+
+
+def openai_chat_tool_choice(tool_choice: ToolChoice, ) -> ChatCompletionToolChoiceOptionParam:
+    if isinstance(tool_choice, ToolFunction):
+        return ChatCompletionNamedToolChoiceParam(type='function', function=dict(name=tool_choice.name))
+    # openai supports 'any' via the 'required' keyword
+    elif tool_choice == 'any':
+        return 'required'
+    else:
+        return tool_choice
+
+
+def chat_tool_calls_from_openai(message: ChatCompletionMessage, tools: List[ToolInfo]) -> Optional[List[ToolCall]]:
+    if message.tool_calls:
+        return [
+            parse_tool_call(call.id, call.function.name, call.function.arguments, tools) for call in message.tool_calls
+        ]
+    else:
+        return None
+
+
+def chat_messages_from_openai(
+    model: str,
+    messages: List[ChatCompletionMessageParam],
+) -> List[ChatMessage]:
+    # track tool names by id
+    tool_names: Dict[str, str] = {}
+
+    chat_messages: List[ChatMessage] = []
+
+    for message in messages:
+        content: Union[str, List[Content]] = []
+        if message['role'] == 'system' or message['role'] == 'developer':
+            sys_content = message['content']
+            if isinstance(sys_content, str):
+                chat_messages.append(ChatMessageSystem(content=sys_content))
+            else:
+                content = []
+                for sc in sys_content:
+                    content.extend(content_from_openai(sc))
+                chat_messages.append(ChatMessageSystem(content=content))
+        elif message['role'] == 'user':
+            user_content = message['content']
+            if isinstance(user_content, str):
+                chat_messages.append(ChatMessageUser(content=user_content))
+            else:
+                content = []
+                for uc in user_content:
+                    content.extend(content_from_openai(uc))
+                chat_messages.append(ChatMessageUser(content=content))
+        elif message['role'] == 'assistant':
+            # resolve content
+            refusal: Optional[Literal[True]] = None
+            internal: Optional[JsonValue] = None
+            asst_content = message.get('content', None)
+            if isinstance(asst_content, str):
+                # Even though the choices API doesn't take advantage of .internal,
+                # we could be transforming from OpenAI choices to Inspect for agent
+                # bridge scenarios where a different model (that does use .internal)
+                # is the actual model being used.
+                asst_content, internal = _parse_content_with_internal(asst_content)
+                asst_content, smuggled_reasoning = parse_content_with_reasoning(asst_content)
+                if smuggled_reasoning:
+                    content = [
+                        smuggled_reasoning,
+                        ContentText(text=asst_content),
+                    ]
+                else:
+                    content = asst_content
+            elif asst_content is None:
+                content = message.get('refusal', None) or ''
+                if content:
+                    refusal = True
+            else:
+                content = []
+                for ac in asst_content:
+                    content.extend(content_from_openai(ac, parse_reasoning=True))
+
+            # resolve reasoning (OpenAI doesn't suport this however OpenAI-compatible
+            # interfaces e.g. DeepSeek do include this field so we pluck it out)
+            reasoning = message.get('reasoning_content', None) or message.get('reasoning', None)
+            if reasoning is not None:
+                # normalize content to an array
+                if isinstance(content, str):
+                    content = [ContentText(text=content, refusal=refusal)]
+
+                # insert reasoning
+                content.insert(0, ContentReasoning(reasoning=str(reasoning)))
+
+            # return message
+            if 'tool_calls' in message:
+                tool_calls: List[ToolCall] = []
+                for call in message['tool_calls']:
+                    tool_calls.append(tool_call_from_openai(call))
+                    tool_names[call['id']] = call['function']['name']
+
+            else:
+                tool_calls = []
+
+            chat_messages.append(
+                ChatMessageAssistant(
+                    content=content,
+                    tool_calls=tool_calls or None,
+                    model=model,
+                    source='generate',
+                    internal=internal,
+                )
+            )
+        elif message['role'] == 'tool':
+            tool_content = message.get('content', None) or ''
+            if isinstance(tool_content, str):
+                # If tool_content is a simple str, it could be the result of some
+                # sub-agent tool call that has <think> or <internal> smuggled inside
+                # of it to support agent bridge scenarios. We have to strip that
+                # data. To be clear, if it's <think>, we'll strip the <think> tag,
+                # but the reasoning summary itself will remain in the content.
+                content, _ = _parse_content_with_internal(tool_content)
+                content, _ = parse_content_with_reasoning(content)
+            else:
+                content = []
+                for tc in tool_content:
+                    content.extend(content_from_openai(tc))
+            chat_messages.append(
+                ChatMessageTool(
+                    content=content,
+                    tool_call_id=message['tool_call_id'],
+                    function=tool_names.get(message['tool_call_id'], ''),
+                )
+            )
+        else:
+            raise ValueError(f'Unexpected message param type: {type(message)}')
+
+    return chat_messages
+
+
+def tool_call_from_openai(tool_call: ChatCompletionMessageToolCallParam) -> ToolCall:
+    return parse_tool_call(
+        tool_call['id'],
+        tool_call['function']['name'],
+        tool_call['function']['arguments'],
+    )
+
+
+def content_from_openai(
+    content: Union[ChatCompletionContentPartParam, ChatCompletionContentPartRefusalParam],
+    parse_reasoning: bool = False,
+) -> List[Content]:
+    # Some providers omit the type tag and use "object-with-a-single-field" encoding
+    if 'type' not in content and len(content) == 1:
+        content['type'] = list(content.keys())[0]  # type: ignore[arg-type]
+    if content['type'] == 'text':
+        text = content['text']
+        if parse_reasoning:
+            content_text, content_reasoning = parse_content_with_reasoning(text)
+            if content_reasoning:
+                return [
+                    content_reasoning,
+                    ContentText(text=content_text),
+                ]
+            else:
+                return [ContentText(text=text)]
+        else:
+            return [ContentText(text=text)]
+    elif content['type'] == 'reasoning':  # type: ignore[comparison-overlap]
+        return [ContentReasoning(reasoning=content['reasoning'])]
+    elif content['type'] == 'image_url':
+        return [ContentImage(image=content['image_url']['url'], detail=content['image_url']['detail'])]
+    elif content['type'] == 'input_audio':
+        return [ContentAudio(
+            audio=content['input_audio']['data'],
+            format=content['input_audio']['format'],
+        )]
+    elif content['type'] == 'refusal':
+        return [ContentText(text=content['refusal'], refusal=True)]
+    else:
+        content_type = content['type']
+        raise ValueError(f"Unexpected content type '{content_type}' in message.")
+
+
+def chat_message_assistant_from_openai(
+    model: str, message: ChatCompletionMessage, tools: List[ToolInfo]
+) -> ChatMessageAssistant:
+    refusal = getattr(message, 'refusal', None)
+    reasoning = getattr(message, 'reasoning_content', None) or getattr(message, 'reasoning', None)
+
+    msg_content = refusal or message.content or ''
+    if reasoning is not None:
+        content: Union[str, List[Content]] = [
+            ContentReasoning(reasoning=str(reasoning)),
+            ContentText(text=msg_content, refusal=True if refusal else None),
+        ]
+    elif refusal is not None:
+        content = [ContentText(text=msg_content, refusal=True)]
+    else:
+        content = msg_content
+
+    return ChatMessageAssistant(
+        content=content,
+        model=model,
+        source='generate',
+        tool_calls=chat_tool_calls_from_openai(message, tools),
+    )
+
+
+def model_output_from_openai(
+    completion: ChatCompletion,
+    choices: list[ChatCompletionChoice],
+) -> ModelOutput:
+    return ModelOutput(
+        model=completion.model,
+        choices=choices,
+        usage=(
+            ModelUsage(
+                input_tokens=completion.usage.prompt_tokens,
+                output_tokens=completion.usage.completion_tokens,
+                input_tokens_cache_read=(
+                    completion.usage.prompt_tokens_details.cached_tokens if completion.usage.prompt_tokens_details
+                    is not None else None  # openai only have cache read stats/pricing.
+                ),
+                reasoning_tokens=(
+                    completion.usage.completion_tokens_details.reasoning_tokens
+                    if completion.usage.completion_tokens_details is not None else None
+                ),
+                total_tokens=completion.usage.total_tokens,
+            ) if completion.usage else None
+        ),
+    )
+
+
+def chat_choices_from_openai(response: ChatCompletion, tools: List[ToolInfo]) -> List[ChatCompletionChoice]:
+    choices = list(response.choices)
+    choices.sort(key=lambda c: c.index)
+    return [
+        ChatCompletionChoice(
+            message=chat_message_assistant_from_openai(response.model, choice.message, tools),
+            stop_reason=as_stop_reason(choice.finish_reason),
+            logprobs=(
+                Logprobs(**choice.logprobs.model_dump())
+                if choice.logprobs and choice.logprobs.content is not None else None
+            ),
+        ) for choice in choices
+    ]
+
+
+def openai_handle_bad_request(model_name: str, e: APIStatusError) -> Union[ModelOutput, Exception]:
+    # extract message
+    if isinstance(e.body, dict) and 'message' in e.body.keys():
+        content = str(e.body.get('message'))
+    else:
+        content = e.message
+
+    # narrow stop_reason
+    stop_reason: Optional[StopReason] = None
+    if e.code == 'context_length_exceeded':
+        stop_reason = 'model_length'
+    elif (
+        e.code == 'invalid_prompt'  # seems to happen for o1/o3
+        or e.code == 'content_policy_violation'  # seems to happen for vision
+        or e.code == 'content_filter'  # seems to happen on azure
+    ):
+        stop_reason = 'content_filter'
+
+    if stop_reason:
+        return ModelOutput.from_content(model=model_name, content=content, stop_reason=stop_reason)
+    else:
+        raise e
+
+
+def openai_media_filter(key: Optional[JsonValue], value: JsonValue) -> JsonValue:
+    # remove images from raw api call
+    if key == 'output' and isinstance(value, dict) and 'image_url' in value:
+        value = copy(value)
+        value.update(image_url=BASE_64_DATA_REMOVED)
+    if key == 'image_url' and isinstance(value, dict) and 'url' in value:
+        url = str(value.get('url'))
+        if url.startswith('data:'):
+            value = copy(value)
+            value.update(url=BASE_64_DATA_REMOVED)
+    elif key == 'input_audio' and isinstance(value, dict) and 'data' in value:
+        value = copy(value)
+        value.update(data=BASE_64_DATA_REMOVED)
+    return value
+
+
+def _parse_content_with_internal(content: str, ) -> Tuple[str, Optional[JsonValue]]:
+    """
+    Extracts and removes a smuggled <internal>...</internal> tag from the content string, if present.
+
+    Note:
+        This OpenAI model does not natively use `.internal`. However, in bridge
+        scenarios—where output from a model that does use `.internal` is routed
+        through this code—such a tag may be present and should be handled.
+
+    Args:
+        content: The input string, possibly containing an <internal> tag with
+        base64-encoded JSON.
+
+    Returns:
+        tuple[str, JsonValue | None]:
+            - The content string with the <internal>...</internal> tag removed (if present), otherwise the original string.
+            - The decoded and parsed internal value (if present), otherwise None.
+
+    Raises:
+        json.JSONDecodeError: If the content of the <internal> tag is not valid JSON after decoding.
+        UnicodeDecodeError: If the content of the <internal> tag is not valid UTF-8 after base64 decoding.
+    """  # noqa: E501
+    internal_pattern = r'<internal>(.*?)</internal>'
+    internal_match = re.search(r'<internal>(.*?)</internal>', content, re.DOTALL)
+
+    return ((
+        re.sub(internal_pattern, '', content, flags=re.DOTALL).strip(),
+        json.loads(base64.b64decode(internal_match.group(1)).decode('utf-8')),
+    ) if internal_match else (content, None))
+
+
+def collect_stream_response(response_stream: List[ChatCompletionChunk]) -> ChatCompletion:
+    collected_chunks: List[ChatCompletionChunk] = []
+    collected_messages = defaultdict(list)
+    collected_reasoning = defaultdict(list)
+    collected_tool_calls = defaultdict(dict)
+
+    for chunk in response_stream:
+        collected_chunks.append(chunk)
+        for choice in chunk.choices:
+            # Handle reasoning content
+            if hasattr(choice.delta, 'reasoning_content') and choice.delta.reasoning_content is not None:
+                collected_reasoning[choice.index].append(choice.delta.reasoning_content)
+
+            # Handle regular content
+            if choice.delta.content is not None:
+                collected_messages[choice.index].append(choice.delta.content)
+
+            # Handle tool calls
+            if hasattr(choice.delta, 'tool_calls') and choice.delta.tool_calls:
+                for tool_call in choice.delta.tool_calls:
+                    tool_id = tool_call.index
+
+                    # Initialize tool call if not present
+                    if tool_id not in collected_tool_calls[choice.index]:
+                        collected_tool_calls[choice.index][tool_id] = {
+                            'id': tool_call.id if hasattr(tool_call, 'id') and tool_call.id else None,
+                            'type': tool_call.type if hasattr(tool_call, 'type') and tool_call.type else None,
+                            'function': {
+                                'name': '',
+                                'arguments': ''
+                            }
+                        }
+
+                    # Update tool call with new chunks
+                    if hasattr(tool_call, 'function'):
+                        if hasattr(tool_call.function, 'name') and tool_call.function.name:
+                            collected_tool_calls[choice.index][tool_id]['function']['name'] = tool_call.function.name
+
+                        if hasattr(tool_call.function, 'arguments') and tool_call.function.arguments:
+                            collected_tool_calls[choice.index
+                                                 ][tool_id]['function']['arguments'] += tool_call.function.arguments
+
+                    # Update ID if it was received later
+                    if hasattr(tool_call, 'id') and tool_call.id:
+                        collected_tool_calls[choice.index][tool_id]['id'] = tool_call.id
+
+    # Get all unique choice indices from all collections
+    all_indices = set(collected_messages.keys()) | set(collected_reasoning.keys()) | set(collected_tool_calls.keys())
+
+    choices = []
+    for index in all_indices:
+        full_reply_content = ''.join(collected_messages.get(index, []))
+        reasoning = ''.join(collected_reasoning.get(index, []))
+
+        # Process tool_calls for this choice if any exists
+        tool_calls_list = None
+        if index in collected_tool_calls and collected_tool_calls[index]:
+            tool_calls_list = list(collected_tool_calls[index].values())
+            # Filter out any tool calls with None id (incomplete tool calls)
+            tool_calls_list = [tc for tc in tool_calls_list if tc['id'] is not None]
+
+        # use the finish_reason from the last chunk that generated this choice
+        finish_reason = None
+        for chunk in reversed(collected_chunks):
+            if chunk.choices and chunk.choices[0].index == index:
+                finish_reason = chunk.choices[0].finish_reason
+                break
+
+        message_kwargs = {'role': 'assistant', 'content': full_reply_content}
+
+        if reasoning:
+            message_kwargs['reasoning_content'] = reasoning
+
+        if tool_calls_list:
+            message_kwargs['tool_calls'] = tool_calls_list
+
+        choice = Choice(
+            finish_reason=finish_reason or 'stop', index=index, message=ChatCompletionMessage(**message_kwargs)
+        )
+        choices.append(choice)
+
+    # build the final completion object
+    return ChatCompletion(
+        id=collected_chunks[0].id,
+        choices=choices,
+        created=collected_chunks[0].created,
+        model=collected_chunks[0].model,
+        object='chat.completion',
+        usage=collected_chunks[-1].usage  # use the usage from the last chunk
+    )