570 lines
21 KiB
Python
570 lines
21 KiB
Python
"""
|
|
OpenAI 提供商适配器
|
|
"""
|
|
import logging
|
|
from typing import AsyncGenerator, Dict, List, Any
|
|
|
|
from .base import BaseLLMProvider, LLMConfig, LLMMessage, LLMResponse, LLMStreamChunk, ToolCall
|
|
from .registry import ProviderRegistry
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
@ProviderRegistry.register
|
|
class OpenAIProvider(BaseLLMProvider):
|
|
"""
|
|
OpenAI 提供商适配器
|
|
|
|
支持 GPT-3.5, GPT-4 等模型
|
|
"""
|
|
|
|
provider_type = 'openai'
|
|
provider_name = 'OpenAI'
|
|
supported_model_types = ['chat', 'completion', 'embedding']
|
|
|
|
DEFAULT_API_BASE = 'https://api.openai.com/v1'
|
|
|
|
def __init__(self, api_key: str = '', api_base: str = '', **kwargs):
|
|
super().__init__(api_key, api_base, **kwargs)
|
|
self.api_base = api_base or self.DEFAULT_API_BASE
|
|
self._client = None
|
|
self._async_client = None
|
|
|
|
def _get_client(self):
|
|
"""获取同步客户端"""
|
|
if self._client is None:
|
|
try:
|
|
from openai import OpenAI
|
|
self._client = OpenAI(
|
|
api_key=self.api_key,
|
|
base_url=self.api_base,
|
|
)
|
|
except ImportError:
|
|
raise ImportError('请安装 openai 库: pip install openai')
|
|
return self._client
|
|
|
|
def _get_async_client(self):
|
|
"""获取异步客户端"""
|
|
if self._async_client is None:
|
|
try:
|
|
from openai import AsyncOpenAI
|
|
self._async_client = AsyncOpenAI(
|
|
api_key=self.api_key,
|
|
base_url=self.api_base,
|
|
)
|
|
except ImportError:
|
|
raise ImportError('请安装 openai 库: pip install openai')
|
|
return self._async_client
|
|
|
|
def chat(
|
|
self,
|
|
messages: List[LLMMessage],
|
|
config: LLMConfig,
|
|
) -> LLMResponse:
|
|
"""同步对话"""
|
|
client = self._get_client()
|
|
|
|
# 构建请求参数
|
|
kwargs = {
|
|
'model': config.model,
|
|
'messages': [m.to_dict() for m in messages],
|
|
'temperature': config.temperature,
|
|
'top_p': config.top_p,
|
|
'max_tokens': config.max_tokens,
|
|
'stop': config.stop,
|
|
'presence_penalty': config.presence_penalty,
|
|
'frequency_penalty': config.frequency_penalty,
|
|
**config.extra_params,
|
|
}
|
|
|
|
# 添加 Function Calling 参数
|
|
if config.tools:
|
|
kwargs['tools'] = [t.to_openai_format() for t in config.tools]
|
|
if config.tool_choice == 'required':
|
|
kwargs['tool_choice'] = 'required'
|
|
elif config.tool_choice == 'none':
|
|
kwargs['tool_choice'] = 'none'
|
|
elif config.tool_choice != 'auto':
|
|
# 指定具体工具
|
|
kwargs['tool_choice'] = {'type': 'function', 'function': {'name': config.tool_choice}}
|
|
else:
|
|
kwargs['tool_choice'] = 'auto'
|
|
|
|
response = client.chat.completions.create(**kwargs)
|
|
|
|
choice = response.choices[0]
|
|
usage = response.usage
|
|
|
|
# 解析工具调用
|
|
tool_calls = None
|
|
if choice.message.tool_calls:
|
|
tool_calls = self._parse_tool_calls(choice.message.tool_calls)
|
|
|
|
return LLMResponse(
|
|
content=choice.message.content or '',
|
|
model=response.model,
|
|
prompt_tokens=usage.prompt_tokens if usage else 0,
|
|
completion_tokens=usage.completion_tokens if usage else 0,
|
|
total_tokens=usage.total_tokens if usage else 0,
|
|
finish_reason=choice.finish_reason or '',
|
|
raw_response=response.model_dump(),
|
|
tool_calls=tool_calls,
|
|
)
|
|
|
|
def _parse_tool_calls(self, tool_calls) -> List[ToolCall]:
|
|
"""解析工具调用"""
|
|
import json
|
|
result = []
|
|
for tc in tool_calls:
|
|
try:
|
|
arguments = json.loads(tc.function.arguments) if tc.function.arguments else {}
|
|
except json.JSONDecodeError:
|
|
arguments = {'raw': tc.function.arguments}
|
|
|
|
result.append(ToolCall(
|
|
id=tc.id,
|
|
name=tc.function.name,
|
|
arguments=arguments,
|
|
))
|
|
return result
|
|
|
|
async def chat_async(
|
|
self,
|
|
messages: List[LLMMessage],
|
|
config: LLMConfig,
|
|
) -> LLMResponse:
|
|
"""异步对话"""
|
|
client = self._get_async_client()
|
|
|
|
# 构建请求参数
|
|
kwargs = {
|
|
'model': config.model,
|
|
'messages': [m.to_dict() for m in messages],
|
|
'temperature': config.temperature,
|
|
'top_p': config.top_p,
|
|
'max_tokens': config.max_tokens,
|
|
'stop': config.stop,
|
|
'presence_penalty': config.presence_penalty,
|
|
'frequency_penalty': config.frequency_penalty,
|
|
**config.extra_params,
|
|
}
|
|
|
|
# 添加 Function Calling 参数
|
|
if config.tools:
|
|
kwargs['tools'] = [t.to_openai_format() for t in config.tools]
|
|
if config.tool_choice == 'required':
|
|
kwargs['tool_choice'] = 'required'
|
|
elif config.tool_choice == 'none':
|
|
kwargs['tool_choice'] = 'none'
|
|
elif config.tool_choice != 'auto':
|
|
kwargs['tool_choice'] = {'type': 'function', 'function': {'name': config.tool_choice}}
|
|
else:
|
|
kwargs['tool_choice'] = 'auto'
|
|
|
|
response = await client.chat.completions.create(**kwargs)
|
|
|
|
choice = response.choices[0]
|
|
usage = response.usage
|
|
|
|
# 解析工具调用
|
|
tool_calls = None
|
|
if choice.message.tool_calls:
|
|
tool_calls = self._parse_tool_calls(choice.message.tool_calls)
|
|
|
|
return LLMResponse(
|
|
content=choice.message.content or '',
|
|
model=response.model,
|
|
prompt_tokens=usage.prompt_tokens if usage else 0,
|
|
completion_tokens=usage.completion_tokens if usage else 0,
|
|
total_tokens=usage.total_tokens if usage else 0,
|
|
finish_reason=choice.finish_reason or '',
|
|
raw_response=response.model_dump(),
|
|
tool_calls=tool_calls,
|
|
)
|
|
|
|
async def chat_stream(
|
|
self,
|
|
messages: List[LLMMessage],
|
|
config: LLMConfig,
|
|
) -> AsyncGenerator[LLMStreamChunk, None]:
|
|
"""流式对话"""
|
|
client = self._get_async_client()
|
|
|
|
# 构建请求参数
|
|
kwargs = {
|
|
'model': config.model,
|
|
'messages': [m.to_dict() for m in messages],
|
|
'temperature': config.temperature,
|
|
'top_p': config.top_p,
|
|
'max_tokens': config.max_tokens,
|
|
'stop': config.stop,
|
|
'presence_penalty': config.presence_penalty,
|
|
'frequency_penalty': config.frequency_penalty,
|
|
'stream': True,
|
|
'stream_options': {"include_usage": True},
|
|
**config.extra_params,
|
|
}
|
|
|
|
# 添加 Function Calling 参数
|
|
if config.tools:
|
|
kwargs['tools'] = [t.to_openai_format() for t in config.tools]
|
|
if config.tool_choice == 'required':
|
|
kwargs['tool_choice'] = 'required'
|
|
elif config.tool_choice == 'none':
|
|
kwargs['tool_choice'] = 'none'
|
|
elif config.tool_choice != 'auto':
|
|
kwargs['tool_choice'] = {'type': 'function', 'function': {'name': config.tool_choice}}
|
|
else:
|
|
kwargs['tool_choice'] = 'auto'
|
|
|
|
stream = await client.chat.completions.create(**kwargs)
|
|
|
|
# 用于累积工具调用
|
|
tool_call_accumulator = {} # id -> {name, arguments}
|
|
|
|
async for chunk in stream:
|
|
if chunk.choices:
|
|
choice = chunk.choices[0]
|
|
delta = choice.delta
|
|
|
|
# 处理工具调用增量
|
|
tool_call_delta = None
|
|
if hasattr(delta, 'tool_calls') and delta.tool_calls:
|
|
for tc_delta in delta.tool_calls:
|
|
tc_index = tc_delta.index
|
|
if tc_index not in tool_call_accumulator:
|
|
tool_call_accumulator[tc_index] = {
|
|
'id': tc_delta.id or '',
|
|
'name': '',
|
|
'arguments': '',
|
|
}
|
|
if tc_delta.id:
|
|
tool_call_accumulator[tc_index]['id'] = tc_delta.id
|
|
if tc_delta.function:
|
|
if tc_delta.function.name:
|
|
tool_call_accumulator[tc_index]['name'] = tc_delta.function.name
|
|
if tc_delta.function.arguments:
|
|
tool_call_accumulator[tc_index]['arguments'] += tc_delta.function.arguments
|
|
tool_call_delta = tool_call_accumulator[tc_index]
|
|
|
|
# 检查是否完成（包含工具调用）
|
|
tool_calls = None
|
|
if choice.finish_reason == 'tool_calls' or (choice.finish_reason and tool_call_accumulator):
|
|
tool_calls = self._parse_accumulated_tool_calls(tool_call_accumulator)
|
|
|
|
yield LLMStreamChunk(
|
|
content=delta.content or '',
|
|
is_finished=choice.finish_reason is not None,
|
|
finish_reason=choice.finish_reason or '',
|
|
tool_calls=tool_calls,
|
|
tool_call_delta=tool_call_delta,
|
|
)
|
|
|
|
# 最后一个 chunk 包含 usage 信息
|
|
if chunk.usage:
|
|
yield LLMStreamChunk(
|
|
content='',
|
|
is_finished=True,
|
|
prompt_tokens=chunk.usage.prompt_tokens,
|
|
completion_tokens=chunk.usage.completion_tokens,
|
|
total_tokens=chunk.usage.total_tokens,
|
|
)
|
|
|
|
def _parse_accumulated_tool_calls(self, accumulator: Dict) -> List[ToolCall]:
|
|
"""解析累积的工具调用"""
|
|
import json
|
|
result = []
|
|
for idx in sorted(accumulator.keys()):
|
|
tc = accumulator[idx]
|
|
try:
|
|
arguments = json.loads(tc['arguments']) if tc['arguments'] else {}
|
|
except json.JSONDecodeError:
|
|
arguments = {'raw': tc['arguments']}
|
|
|
|
result.append(ToolCall(
|
|
id=tc['id'],
|
|
name=tc['name'],
|
|
arguments=arguments,
|
|
))
|
|
return result
|
|
|
|
def chat_stream_sync(
|
|
self,
|
|
messages: List[LLMMessage],
|
|
config: LLMConfig,
|
|
):
|
|
"""同步流式对话"""
|
|
client = self._get_client()
|
|
|
|
# 构建请求参数
|
|
kwargs = {
|
|
'model': config.model,
|
|
'messages': [m.to_dict() for m in messages],
|
|
'temperature': config.temperature,
|
|
'top_p': config.top_p,
|
|
'max_tokens': config.max_tokens,
|
|
'stop': config.stop,
|
|
'presence_penalty': config.presence_penalty,
|
|
'frequency_penalty': config.frequency_penalty,
|
|
'stream': True,
|
|
'stream_options': {"include_usage": True},
|
|
**config.extra_params,
|
|
}
|
|
|
|
# 添加 Function Calling 参数
|
|
if config.tools:
|
|
kwargs['tools'] = [t.to_openai_format() for t in config.tools]
|
|
if config.tool_choice == 'required':
|
|
kwargs['tool_choice'] = 'required'
|
|
elif config.tool_choice == 'none':
|
|
kwargs['tool_choice'] = 'none'
|
|
elif config.tool_choice != 'auto':
|
|
kwargs['tool_choice'] = {'type': 'function', 'function': {'name': config.tool_choice}}
|
|
else:
|
|
kwargs['tool_choice'] = 'auto'
|
|
|
|
stream = client.chat.completions.create(**kwargs)
|
|
|
|
# 用于累积工具调用
|
|
tool_call_accumulator = {}
|
|
|
|
for chunk in stream:
|
|
if chunk.choices:
|
|
choice = chunk.choices[0]
|
|
delta = choice.delta
|
|
|
|
# 处理工具调用增量
|
|
tool_call_delta = None
|
|
if hasattr(delta, 'tool_calls') and delta.tool_calls:
|
|
for tc_delta in delta.tool_calls:
|
|
tc_index = tc_delta.index
|
|
if tc_index not in tool_call_accumulator:
|
|
tool_call_accumulator[tc_index] = {
|
|
'id': tc_delta.id or '',
|
|
'name': '',
|
|
'arguments': '',
|
|
}
|
|
if tc_delta.id:
|
|
tool_call_accumulator[tc_index]['id'] = tc_delta.id
|
|
if tc_delta.function:
|
|
if tc_delta.function.name:
|
|
tool_call_accumulator[tc_index]['name'] = tc_delta.function.name
|
|
if tc_delta.function.arguments:
|
|
tool_call_accumulator[tc_index]['arguments'] += tc_delta.function.arguments
|
|
tool_call_delta = tool_call_accumulator[tc_index]
|
|
|
|
# 检查是否完成（包含工具调用）
|
|
tool_calls = None
|
|
if choice.finish_reason == 'tool_calls' or (choice.finish_reason and tool_call_accumulator):
|
|
tool_calls = self._parse_accumulated_tool_calls(tool_call_accumulator)
|
|
|
|
yield LLMStreamChunk(
|
|
content=delta.content or '',
|
|
is_finished=choice.finish_reason is not None,
|
|
finish_reason=choice.finish_reason or '',
|
|
tool_calls=tool_calls,
|
|
tool_call_delta=tool_call_delta,
|
|
)
|
|
|
|
# 最后一个 chunk 包含 usage 信息
|
|
if chunk.usage:
|
|
yield LLMStreamChunk(
|
|
content='',
|
|
is_finished=True,
|
|
prompt_tokens=chunk.usage.prompt_tokens,
|
|
completion_tokens=chunk.usage.completion_tokens,
|
|
total_tokens=chunk.usage.total_tokens,
|
|
)
|
|
|
|
# 模型类型推断关键词
|
|
_EMBEDDING_KEYWORDS = ('embed', 'bge', 'nomic-embed', 'mxbai-embed', 'jina-embedding')
|
|
_RERANK_KEYWORDS = ('rerank', 'reranker')
|
|
_SKIP_KEYWORDS = ('dall-e', 'tts', 'whisper', 'audio', 'moderation', 'davinci', 'babbage')
|
|
|
|
async def fetch_models_from_api(self) -> List[Dict[str, Any]]:
|
|
"""
|
|
通过 /v1/models API 在线拉取最新模型列表
|
|
|
|
适用于所有 OpenAI 兼容的提供商
|
|
"""
|
|
base_url = (self.api_base or self.DEFAULT_API_BASE).rstrip('/')
|
|
try:
|
|
return await self.fetch_models_from_api_strict()
|
|
except Exception as e:
|
|
logger.warning(f'在线拉取模型列表失败 ({base_url}): {e}')
|
|
return []
|
|
|
|
async def fetch_models_from_api_strict(self) -> List[Dict[str, Any]]:
|
|
"""通过 /models 端点严格拉取模型列表，失败时保留上游异常"""
|
|
import httpx
|
|
|
|
base_url = (self.api_base or self.DEFAULT_API_BASE).rstrip('/')
|
|
headers = {}
|
|
if self.api_key:
|
|
headers['Authorization'] = f'Bearer {self.api_key}'
|
|
|
|
async with httpx.AsyncClient(timeout=15) as client:
|
|
resp = await client.get(f'{base_url}/models', headers=headers)
|
|
resp.raise_for_status()
|
|
data = resp.json()
|
|
|
|
model_list = data.get('data', [])
|
|
if not model_list:
|
|
return []
|
|
|
|
results = []
|
|
for item in model_list:
|
|
model_id = item.get('id', '')
|
|
if not model_id:
|
|
continue
|
|
|
|
model_lower = model_id.lower()
|
|
|
|
# 跳过非文本模型
|
|
if any(kw in model_lower for kw in self._SKIP_KEYWORDS):
|
|
continue
|
|
|
|
# 推断模型类型
|
|
if any(kw in model_lower for kw in self._RERANK_KEYWORDS):
|
|
model_type = 'rerank'
|
|
elif any(kw in model_lower for kw in self._EMBEDDING_KEYWORDS):
|
|
model_type = 'embedding'
|
|
else:
|
|
model_type = 'chat'
|
|
|
|
results.append({
|
|
'model_name': model_id,
|
|
'display_name': model_id,
|
|
'model_type': model_type,
|
|
'max_tokens': 4096,
|
|
'context_window': 4096,
|
|
'supports_vision': False,
|
|
'supports_function_call': model_type == 'chat',
|
|
'input_price': 0,
|
|
'output_price': 0,
|
|
})
|
|
|
|
# 补充默认列表中的 embedding/rerank 模型（API 可能不返回这些类型）
|
|
fetched_names = {m['model_name'] for m in results}
|
|
for dm in self.__class__.get_default_models():
|
|
if dm['model_type'] in ('embedding', 'rerank') and dm['model_name'] not in fetched_names:
|
|
results.append(dm)
|
|
|
|
# 按类型排序：chat → embedding → rerank
|
|
type_order = {'chat': 0, 'embedding': 1, 'rerank': 2}
|
|
results.sort(key=lambda m: (type_order.get(m['model_type'], 9), m['model_name']))
|
|
logger.info(f'从 {base_url} 拉取到 {len(results)} 个模型')
|
|
return results
|
|
|
|
@classmethod
|
|
def get_default_models(cls) -> List[Dict[str, Any]]:
|
|
"""获取默认模型列表"""
|
|
return [
|
|
# ---- Chat 模型 ----
|
|
{
|
|
'model_name': 'o4-mini',
|
|
'display_name': 'o4-mini',
|
|
'model_type': 'chat',
|
|
'max_tokens': 16384,
|
|
'context_window': 200000,
|
|
'supports_vision': True,
|
|
'supports_function_call': True,
|
|
'input_price': 0.0011,
|
|
'output_price': 0.0044,
|
|
},
|
|
{
|
|
'model_name': 'o3-mini',
|
|
'display_name': 'o3-mini',
|
|
'model_type': 'chat',
|
|
'max_tokens': 16384,
|
|
'context_window': 200000,
|
|
'supports_vision': False,
|
|
'supports_function_call': True,
|
|
'input_price': 0.0011,
|
|
'output_price': 0.0044,
|
|
},
|
|
{
|
|
'model_name': 'gpt-4.1',
|
|
'display_name': 'GPT-4.1',
|
|
'model_type': 'chat',
|
|
'max_tokens': 32768,
|
|
'context_window': 1047576,
|
|
'supports_vision': True,
|
|
'supports_function_call': True,
|
|
'input_price': 0.002,
|
|
'output_price': 0.008,
|
|
},
|
|
{
|
|
'model_name': 'gpt-4.1-mini',
|
|
'display_name': 'GPT-4.1 Mini',
|
|
'model_type': 'chat',
|
|
'max_tokens': 32768,
|
|
'context_window': 1047576,
|
|
'supports_vision': True,
|
|
'supports_function_call': True,
|
|
'input_price': 0.0004,
|
|
'output_price': 0.0016,
|
|
},
|
|
{
|
|
'model_name': 'gpt-4.1-nano',
|
|
'display_name': 'GPT-4.1 Nano',
|
|
'model_type': 'chat',
|
|
'max_tokens': 32768,
|
|
'context_window': 1047576,
|
|
'supports_vision': True,
|
|
'supports_function_call': True,
|
|
'input_price': 0.0001,
|
|
'output_price': 0.0004,
|
|
},
|
|
{
|
|
'model_name': 'gpt-4o',
|
|
'display_name': 'GPT-4o',
|
|
'model_type': 'chat',
|
|
'max_tokens': 16384,
|
|
'context_window': 128000,
|
|
'supports_vision': True,
|
|
'supports_function_call': True,
|
|
'input_price': 0.0025,
|
|
'output_price': 0.01,
|
|
},
|
|
{
|
|
'model_name': 'gpt-4o-mini',
|
|
'display_name': 'GPT-4o Mini',
|
|
'model_type': 'chat',
|
|
'max_tokens': 16384,
|
|
'context_window': 128000,
|
|
'supports_vision': True,
|
|
'supports_function_call': True,
|
|
'input_price': 0.00015,
|
|
'output_price': 0.0006,
|
|
},
|
|
# ---- Embedding 模型 ----
|
|
{
|
|
'model_name': 'text-embedding-3-small',
|
|
'display_name': 'Text Embedding 3 Small',
|
|
'model_type': 'embedding',
|
|
'max_tokens': 8191,
|
|
'context_window': 8191,
|
|
'input_price': 0.00002,
|
|
'output_price': 0,
|
|
},
|
|
{
|
|
'model_name': 'text-embedding-3-large',
|
|
'display_name': 'Text Embedding 3 Large',
|
|
'model_type': 'embedding',
|
|
'max_tokens': 8191,
|
|
'context_window': 8191,
|
|
'input_price': 0.00013,
|
|
'output_price': 0,
|
|
},
|
|
{
|
|
'model_name': 'text-embedding-ada-002',
|
|
'display_name': 'Text Embedding Ada 002',
|
|
'model_type': 'embedding',
|
|
'max_tokens': 8191,
|
|
'context_window': 8191,
|
|
'input_price': 0.0001,
|
|
'output_price': 0,
|
|
},
|
|
]
|