嵌入模型api接口支持 #1463
嵌入模型api接口支持 #1463
Conversation
重新排序请求的序列图 (Voyage)sequenceDiagram
participant Client
participant VoyageAdapter
participant VoyageAPI as Voyage Rerank API
Client->>+VoyageAdapter: rerank(LLMReRankRequest)
VoyageAdapter->>+VoyageAPI: POST /v1/rerank
VoyageAPI-->>-VoyageAdapter: RerankResponse (JSON)
VoyageAdapter->>VoyageAdapter: Process response, sort results
VoyageAdapter-->>-Client: LLMReRankResponse
文件级别更改
提示和命令与 Sourcery 互动
自定义您的体验访问您的 仪表板 以：
获得帮助Original review guide in EnglishReviewer's GuideThis pull request introduces support for embedding models by defining standard request/response formats (LLMEmbeddingRequest, LLMEmbeddingResponse) and an LLMEmbeddingProtocol. This protocol is implemented in the existing OpenAI, Gemini, and Ollama adapters, as well as a newly added Voyage adapter, by calling the respective provider APIs. The Voyage adapter also introduces reranking capabilities via a new LLMReRankProtocol and corresponding formats. Base adapter classes and protocols were refactored to accommodate these changes, and the minimum Python version was raised to 3.10. Sequence Diagram for Embedding RequestsequenceDiagram
participant Client
participant Adapter as LLMAdapter (e.g., VoyageAdapter)
participant ProviderAPI as External Provider API (e.g., Voyage API)
Client->>+Adapter: embed(LLMEmbeddingRequest)
Adapter->>+ProviderAPI: POST /v1/embeddings (or similar)
ProviderAPI-->>-Adapter: EmbeddingResponse (JSON)
Adapter->>Adapter: Process response
Adapter-->>-Client: LLMEmbeddingResponse
Sequence Diagram for Reranking Request (Voyage)sequenceDiagram
participant Client
participant VoyageAdapter
participant VoyageAPI as Voyage Rerank API
Client->>+VoyageAdapter: rerank(LLMReRankRequest)
VoyageAdapter->>+VoyageAPI: POST /v1/rerank
VoyageAPI-->>-VoyageAdapter: RerankResponse (JSON)
VoyageAdapter->>VoyageAdapter: Process response, sort results
VoyageAdapter-->>-Client: LLMReRankResponse
File-Level Changes
Tips and commandsInteracting with Sourcery
Customizing Your ExperienceAccess your dashboard to:
Getting Help
|
|
有一点点问题，两个分支代码混淆了等待我弄好再说 |
Kaos-dawn
force-pushed
the
feature/embed
branch
from
a804a5b to
a1ca514
Compare
Kaos-dawn
force-pushed
the
feature/embed
branch
from
446a500 to
75bac27
Compare
refactor: 优化了llm_adapter使其使用组合接口实现基本的功能。 feature: 添加了rerank接口以实现重排器模型。并设计了对应的模型。 feature: 添加了embedding接口以实现embedding模型。并设计了对应的模型。 feature: 为主要模型添加了其嵌入式api接口。 feature: 添加了一个嵌入模型以及重排器模型提供商。（暂时未注册） merged: 合并了主线mc相关内容并解决了对应冲突 fix: 尝试修复pytest失败用例 fix: 修复mypy检查出的问题
Kaos-dawn
force-pushed
the
feature/embed
branch
from
75bac27 to
f86ac92
Compare
| super().__init__(config) | ||
|
|
||
| def embed(self, req): | ||
| raise NotImplementedError("Minimax does not support embedding") |
| } | ||
| data = { | ||
| "model": req.model, | ||
| "inputs": asyncio.run(resolve_media_base64(req.inputs, media_manager)), |
There was a problem hiding this comment.
有模拟的mediaManager吗我去写个pytest试试
|
|
||
|
|
||
| class OpenAIAdapter(LLMBackendAdapter, AutoDetectModelsProtocol): | ||
| class OpenAIAdapter(LLMBackendAdapter, AutoDetectModelsProtocol, LLMChatProtocol, LLMEmbeddingProtocol): |
There was a problem hiding this comment.
既然此处的 OpenAI Adapter 实现了 Embedding Protocol 已不能被其他 adapter 复用，建议抽离 chat 能力为单独的类，以便其他继承了 OpenAI Adapter 的 adapter 使用
MyPy 类型检查通过 ✅PR 修改的代码行通过了类型检查。 |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| raise e | ||
|
|
||
| return LLMEmbeddingResponse( | ||
| vectors=[data["embedding"] for data in response_data["data"]], |
There was a problem hiding this comment.
MyPy 类型错误: TypedDict "ReRankData" has no key "embedding" (typeddict-item)
详细信息请参考 mypy 文档。
| raise e | ||
|
|
||
| return LLMEmbeddingResponse( | ||
| vectors=[data["embedding"] for data in response_data["data"]], |
There was a problem hiding this comment.
MyPy 类型错误: TypedDict "ReRankData" has no key "embedding" (typeddict-item)
详细信息请参考 mypy 文档。
| "model": req.model, | ||
| "dimensions": req.dimension, | ||
| "encoding_format": req.encoding_format, | ||
| "user": req.user |
There was a problem hiding this comment.
MyPy 类型错误: "LLMEmbeddingRequest" has no attribute "user" (attr-defined)
详细信息请参考 mypy 文档。
Kaos-dawn
force-pushed
the
feature/embed
branch
from
ed7dd18 to
313a8ae
Compare
| "model": req.model, | ||
| "dimensions": req.dimension, | ||
| "encoding_format": req.encoding_format, | ||
| "user": req.user |
There was a problem hiding this comment.
MyPy 类型错误: "LLMEmbeddingRequest" has no attribute "user" (attr-defined)
详细信息请参考 mypy 文档。
refactor: 改写了voyage 的转化逻辑，考虑到了事件循环导致的问题。 refactor: 移除了LLMEmbeddingRequest中的user字段。 feature: 编写了一个pytest用例用于测试voyage_adapter中resolve_media_base64是否正常运行。 fix: 添加了voyage的嵌入式接口返回值类型标识，修复了错误的类型标识。 fix: 修复了openai 嵌入式接口user字段未清理的问题。 tips: 对于腾讯, 阿里云等提供商，暂时继承全量openai接口。
Kaos-dawn
force-pushed
the
feature/embed
branch
from
313a8ae to
63a1929
Compare
![sourcery-ai[bot]](https://avatars.githubusercontent.com/in/48477?s=60&v=4){kind=small}
There was a problem hiding this comment.
嘿 @teywat - 我已经审查了你的更改 - 这里有一些反馈：
- 新的
embed和rerank方法使用同步网络调用 (requests)，可能会阻塞异步事件循环，这在混合同步/异步调用的VoyageAdapter中尤其明显；考虑始终如一地使用异步 HTTP 客户端。 - 阐明 pyproject.toml 中 Python 版本提升到 3.10 是否是此 PR 中引入的功能严格要求的。
这是我在审查期间查看的内容
- 🟡 一般问题：发现 3 个问题
- 🟢 安全性：一切看起来都很好
- 🟡 测试：发现 2 个问题
- 🟡 复杂性：发现 1 个问题
- 🟢 文档：一切看起来都很好
帮助我更有用！请点击每个评论上的 👍 或 👎，我将使用反馈来改进你的评论。
Original comment in English
Hey @teywat - I've reviewed your changes - here's some feedback:
- The new
embedandrerankmethods use synchronous network calls (requests), potentially blocking async event loops, notably seen inVoyageAdapterwhich mixes sync/async calls; consider using an async HTTP client consistently. - Clarify if the Python version bump to 3.10 in pyproject.toml is strictly required by the features introduced in this PR.
Here's what I looked at during the review
- 🟡 General issues: 3 issues found
- 🟢 Security: all looks good
- 🟡 Testing: 2 issues found
- 🟡 Complexity: 1 issue found
- 🟢 Documentation: all looks good
Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.
| logger = get_logger("OpenAIAdapter") | ||
|
|
||
| async def convert_parts_factory(messages: LLMChatMessage, media_manager: MediaManager) -> list[dict]: | ||
| async def convert_parts_factory(messages: LLMChatMessage, media_manager: MediaManager) -> list | list[dict]: |
There was a problem hiding this comment.
suggestion: 重新审视联合返回类型，以使其更清晰。
list | list[dict] 联合是多余的。使用精确的类型（如 list[dict]）或显式的 Union 来反映预期的结构。
| async def convert_parts_factory(messages: LLMChatMessage, media_manager: MediaManager) -> list | list[dict]: | |
| async def convert_parts_factory(messages: LLMChatMessage, media_manager: MediaManager) -> list[dict]: |
Original comment in English
suggestion: Revisit the union return type for clarity.
The union list | list[dict] is redundant. Use a precise type like list[dict] or an explicit Union to reflect the expected structure.
| async def convert_parts_factory(messages: LLMChatMessage, media_manager: MediaManager) -> list | list[dict]: | |
| async def convert_parts_factory(messages: LLMChatMessage, media_manager: MediaManager) -> list[dict]: |
| try: | ||
| response.raise_for_status() | ||
| response_data: EmbeddingResponse = response.json() | ||
| except Exception as e: | ||
| logger.error(f"Response: {response.text}") | ||
| raise e | ||
|
|
||
| return LLMEmbeddingResponse( | ||
| vectors=[data["embedding"] for data in response_data["data"]], | ||
| usage = Usage( |
There was a problem hiding this comment.
suggestion (bug_risk): 审查在可能正在运行的循环中使用 'run_until_complete' 的情况。
将其转换为异步，或者仅在没有事件循环运行时调用它，因为在正在运行的循环上调用 run_until_complete 会引发 RuntimeError。
Original comment in English
suggestion (bug_risk): Review use of 'run_until_complete' in a potentially running loop.
Convert this to async or only invoke it when no event loop is running, as calling run_until_complete on a running loop raises RuntimeError.
| # 删除 None 字段 | ||
| data = {k: v for k, v in data.items() if v is not None} | ||
| logger.debug(f"Request: {data}") | ||
| response = requests.post(api_url, headers=headers, json=data) |
There was a problem hiding this comment.
suggestion (performance): 同步方法中潜在的阻塞网络调用。
确保阻塞的 requests.post 调用在异步上下文中使用时不会使事件循环停滞——卸载它们或切换到异步 HTTP 客户端。
建议的实现：
import asyncio
# (其他导入) response = await asyncio.to_thread(requests.post, api_url, headers=headers, json=data) await asyncio.to_thread(response.raise_for_status) response_data = await asyncio.to_thread(response.json)确保父函数声明为 async，以便 await 调用有效。如果此函数是从同步代码调用的，请相应地调整调用站点。
Original comment in English
suggestion (performance): Potential blocking network call in a synchronous method.
Ensure the blocking requests.post calls don't stall the event loop when used in async contexts—offload them or switch to an async HTTP client.
Suggested implementation:
import asyncio
# (other imports) response = await asyncio.to_thread(requests.post, api_url, headers=headers, json=data) await asyncio.to_thread(response.raise_for_status) response_data = await asyncio.to_thread(response.json)Ensure that the parent function is declared as async so that the await calls are valid. If this function is called from synchronous code, adjust the call site accordingly.
There was a problem hiding this comment.
鉴于现在的调度器执行流程为同步执行。同一时间只有一次http请求。改为异步意义不大。推荐等异步工作流执行特性完成再使用httpx重写
There was a problem hiding this comment.
理解你的顾虑。目前整个调度器确实是同步执行、并且同一时间只有一次 HTTP 请求，单纯把 embed 改成 asyncio.to_thread 并不会带来太大收益。我们可以暂时保留同步实现，待上层引入真正的异步执行流或者切换到 httpx.AsyncClient 后，再一次性把所有后端请求（包括 embedding）整体迁移到异步。
建议在代码里加个 TODO／Issue 标记这块，后续跟进：
- 支持真正的异步工作流时，将 embed 方法改为 async 并使用
httpx.AsyncClient（或其他异步 HTTP 库）。 - 同步场景下，可考虑给
requests.post增加超时、重试等策略，保证稳定性。
这样既不影响当前功能，又能为后续的异步重构留足空间。
tests/test_voyage.py
Outdated
| class MockMediaManager: | ||
| def get_media(self, media_id: str): | ||
| return MockMedia(media_id) | ||
|
|
||
| class MockMedia: | ||
| def __init__(self, media_id: str): | ||
| self.media_id = media_id | ||
|
|
||
| async def get_base64(self) -> str: | ||
| return f"{self.media_id}:base64_data" |
There was a problem hiding this comment.
suggestion (testing): 考虑测试找不到媒体的边缘情况
添加一个 media_manager.get_media 返回 None 的测试，并验证 resolve_media_base64 是否引发 ValueError。
Original comment in English
suggestion (testing): Consider testing the edge case where media is not found
Add a test where media_manager.get_media returns None and verify that resolve_media_base64 raises a ValueError.
tests/test_voyage.py
Outdated
| def test_resolve_media_base64(inputs, media_manager): | ||
| try: | ||
| loop = asyncio.get_running_loop() | ||
| except RuntimeError: | ||
| loop = asyncio.new_event_loop() | ||
| asyncio.set_event_loop(loop) | ||
|
|
||
| results = loop.run_until_complete(resolve_media_base64(inputs, media_manager)) | ||
| assert results[0] == {'content': [{"type": "text", "text": "hello"}]} | ||
| assert results[1] == {"content": [{"type": "image_base64", "image_base64": "114514:base64_data"}]} |
There was a problem hiding this comment.
issue (testing): 测试套件缺少对新的嵌入和重排序功能的覆盖
请为 OpenAI、Gemini、Ollama 和 Voyage 中的新嵌入方法（涵盖文本和多模态，如果适用）、Voyage 重排序方法以及错误处理（无效输入、API 错误、不支持的功能）添加单元或集成测试。
Original comment in English
issue (testing): Test suite lacks coverage for the new embedding and reranking functionalities
Please add unit or integration tests for the new embed methods in OpenAI, Gemini, Ollama, and Voyage (covering text and multi-modal where applicable), the Voyage rerank method, and error handling (invalid inputs, API errors, unsupported features).
| ) | ||
| ) | ||
|
|
||
| def _multi_modal_embedding(self, req: LLMEmbeddingRequest, media_manager: MediaManager) -> LLMEmbeddingResponse: |
There was a problem hiding this comment.
issue (complexity): 考虑将异步媒体解析与同步嵌入过程分离，并使用 asyncio.run 来桥接异步操作。
可操作的重构建议
主要的复杂性来自于将同步 HTTP 调用与异步媒体处理混合在一起。考虑以下两个重构步骤：
-
将异步媒体解析与嵌入分离：
为多模态嵌入创建一个专用的异步方法。这将事件循环处理与同步逻辑隔离。 -
一致地从同步方法桥接异步：
不要尝试在_multi_modal_embedding内部获取或创建正在运行的循环，而是将异步工作委托给明确定义的异步助手，并使用asyncio.run（或类似方法）在外部边界运行它。
重构示例：
# 提取异步功能
async def async_multi_modal_embedding(
self, req: LLMEmbeddingRequest, media_manager: MediaManager
) -> LLMEmbeddingResponse:
api_url = f"{self.config.api_base}/v1/multimodalembeddings"
headers = {
"Authorization": f"Bearer {self.config.api_key}",
"Content-Type": "application/json",
}
inputs = await resolve_media_base64(req.inputs, media_manager)
data = {
"model": req.model,
"inputs": inputs,
"input_type": req.input_type,
"truncation": req.truncate,
"output_encoding": req.encoding_format,
}
data = {k: v for k, v in data.items() if v is not None}
response = requests.post(api_url, headers=headers, json=data)
response.raise_for_status()
response_data: ModalEmbeddingResponse = response.json()
return LLMEmbeddingResponse(
vectors=[item["embedding"] for item in response_data["data"]],
usage=Usage(total_tokens=response_data["usage"].get("total_tokens", 0)),
)
# 在您的 embed 方法中，通过 asyncio.run 委托：
def embed(self, req: LLMEmbeddingRequest) -> LLMEmbeddingResponse:
if all(isinstance(input, LLMChatTextContent) for input in req.inputs):
return self._text_embedding(req)
else:
return asyncio.run(self.async_multi_modal_embedding(req, self.media_manager))总结：
这种方法从适配器中删除了嵌套的事件循环管理，隔离了异步问题，并使其更容易维护和扩展代码，而不会混合职责。
Original comment in English
issue (complexity): Consider separating the asynchronous media resolution from the synchronous embedding process and using asyncio.run to bridge the async operations.
Actionable Refactoring Suggestion
The main complexity comes from mixing synchronous HTTP calls with asynchronous media processing. Consider these two refactoring steps:
-
Separate Async Media Resolution from Embedding:
Create a dedicated async method for multi-modal embedding. This isolates event loop handling from your synchronous logic. -
Consistently Bridge Async from Synchronous Methods:
Instead of trying to get or create a running loop inside_multi_modal_embedding, delegate async work to a clearly defined async helper and run it usingasyncio.run(or similar) at the outer boundary.
Example Refactoring:
# Extract async functionality
async def async_multi_modal_embedding(
self, req: LLMEmbeddingRequest, media_manager: MediaManager
) -> LLMEmbeddingResponse:
api_url = f"{self.config.api_base}/v1/multimodalembeddings"
headers = {
"Authorization": f"Bearer {self.config.api_key}",
"Content-Type": "application/json",
}
inputs = await resolve_media_base64(req.inputs, media_manager)
data = {
"model": req.model,
"inputs": inputs,
"input_type": req.input_type,
"truncation": req.truncate,
"output_encoding": req.encoding_format,
}
data = {k: v for k, v in data.items() if v is not None}
response = requests.post(api_url, headers=headers, json=data)
response.raise_for_status()
response_data: ModalEmbeddingResponse = response.json()
return LLMEmbeddingResponse(
vectors=[item["embedding"] for item in response_data["data"]],
usage=Usage(total_tokens=response_data["usage"].get("total_tokens", 0)),
)
# In your embed method, delegate via asyncio.run:
def embed(self, req: LLMEmbeddingRequest) -> LLMEmbeddingResponse:
if all(isinstance(input, LLMChatTextContent) for input in req.inputs):
return self._text_embedding(req)
else:
return asyncio.run(self.async_multi_modal_embedding(req, self.media_manager))Summary:
This approach removes nested event loop management from within the adapter, isolates async concerns, and makes it easier to maintain and extend the code without mixing responsibilities.
|
话说需不需要将所有adapter的api调用方法改为异步的httpx? |
我觉得可以 |
fix: 修复了所有adapter中loop的使用方法， 原方法未实现事件循环的关闭以及仍被引用无法被垃圾回收的潜在资源泄露风险。将原本手动进行的loop生命周期管理，更换为更高层级封装好的asyncio.run()方法。 refactored: 查阅了api reference。暂时更正了gemini_adapter中api_key的转递方法（没有测试条件，展示未测试实际使用。 pytest测试通过） refactored: 查阅了voyage_api_reference。更改了待嵌入文本以及图像的序列化逻辑。Note: 新逻辑需要用到meida.description属性。 feature: 为llm_adapter添加了标准测试集框架，并实现了所有的embed方法测式集。
| async def get_url(self) -> str: | ||
| return "https://example.com/image.png" | ||
|
|
||
| @property |
There was a problem hiding this comment.
MyPy 类型错误: Read-only property cannot override read-write property (misc)
详细信息请参考 mypy 文档。
| raise ValueError(f"Media {input.media_id} not found") | ||
| results.append({ | ||
| "content": [ | ||
| {"type": "text", "text": media.description}, |
There was a problem hiding this comment.
MyPy 类型错误: Dict entry 1 has incompatible type "str": "str | None"; expected "str": "str" (dict-item)
详细信息请参考 mypy 文档。
| usage = Usage( | ||
| total_tokens = response_data["usage"].get("total_tokens", 0) | ||
| ), | ||
| sort = req.sort |
There was a problem hiding this comment.
MyPy 类型错误: Argument "sort" to "LLMReRankResponse" has incompatible type "bool | None"; expected "bool" (arg-type)
详细信息请参考 mypy 文档。
| response = voyage_adapter.embed(req) | ||
| assert isinstance(response, LLMEmbeddingResponse) | ||
| assert response.vectors[0] == REFERENCE_VECTOR | ||
| assert response.usage.total_tokens == 10 |
There was a problem hiding this comment.
MyPy 类型错误: Item "None" of "Usage | None" has no attribute "total_tokens" (union-attr)
详细信息请参考 mypy 文档。
| response = voyage_adapter.embed(req) | ||
| assert isinstance(response, LLMEmbeddingResponse) | ||
| assert response.vectors[0] == REFERENCE_VECTOR | ||
| assert response.usage.total_tokens == 3576 |
There was a problem hiding this comment.
MyPy 类型错误: Item "None" of "Usage | None" has no attribute "total_tokens" (union-attr)
详细信息请参考 mypy 文档。
| def convert_llm_chat_message_to_openai_message(messages: list[LLMChatMessage], media_manager: MediaManager, loop: asyncio.AbstractEventLoop) -> list[dict]: | ||
| results = loop.run_until_complete( | ||
| def convert_llm_chat_message_to_openai_message(messages: list[LLMChatMessage], media_manager: MediaManager) -> list[dict]: | ||
| results = asyncio.run( |
There was a problem hiding this comment.
MyPy 类型错误: Need type annotation for "results" (var-annotated)
详细信息请参考 mypy 文档。
|
|
||
| data = { | ||
| "contents": loop.run_until_complete(asyncio.gather(*[convert_llm_chat_message_to_gemini_message(msg, self.media_manager) for msg in req.messages])), | ||
| "contents": asyncio.run(asyncio.gather(*[convert_llm_chat_message_to_gemini_message(msg, self.media_manager) for msg in req.messages])), |
There was a problem hiding this comment.
MyPy 类型错误: Argument 1 to "run" has incompatible type "Future[list[dict[Any, Any]]]"; expected "Coroutine[Any, Any, Never]" (arg-type)
详细信息请参考 mypy 文档。
| class TestOllamaAdapter: | ||
| @pytest.fixture(scope="class") | ||
| def ollama_adapter(self, mock_media_manager) -> OllamaAdapter: | ||
| config = OllamaConfig( |
There was a problem hiding this comment.
MyPy 类型错误: Unexpected keyword argument "api_key" for "OllamaConfig" (call-arg)
详细信息请参考 mypy 文档。
fix: 修复openai_adapter中max_tokens参数被移除的问题。 fix: fix mypy提示的类型错误。
| response = voyage_adapter.embed(req) | ||
| assert isinstance(response, LLMEmbeddingResponse) | ||
| assert response.vectors[0] == REFERENCE_VECTOR | ||
| assert response.usage.total_tokens == 10 |
There was a problem hiding this comment.
MyPy 类型错误: Item "None" of "Usage | None" has no attribute "total_tokens" (union-attr)
详细信息请参考 mypy 文档。
|
|
||
| response = openai_adapter.chat(req) | ||
| assert isinstance(response, LLMChatResponse) | ||
| assert response.message.content[0].text == "mock_response" |
There was a problem hiding this comment.
MyPy 类型错误: Item "LLMChatImageContent" of "LLMChatTextContent | LLMChatImageContent | LLMToolCallContent | LLMToolResultContent" has no attribute "text" (union-attr)
详细信息请参考 mypy 文档。
|
|
||
| response = openai_adapter.chat(req) | ||
| assert isinstance(response, LLMChatResponse) | ||
| assert response.message.content[0].text == "mock_response" |
There was a problem hiding this comment.
MyPy 类型错误: Item "LLMToolCallContent" of "LLMChatTextContent | LLMChatImageContent | LLMToolCallContent | LLMToolResultContent" has no attribute "text" (union-attr)
详细信息请参考 mypy 文档。
|
|
||
| response = openai_adapter.chat(req) | ||
| assert isinstance(response, LLMChatResponse) | ||
| assert response.message.content[0].text == "mock_response" |
There was a problem hiding this comment.
MyPy 类型错误: Item "LLMToolResultContent" of "LLMChatTextContent | LLMChatImageContent | LLMToolCallContent | LLMToolResultContent" has no attribute "text" (union-attr)
详细信息请参考 mypy 文档。
| router = APIRouter(tags=["gemini"], dependencies=[Depends(gemini_authenticate)]) | ||
|
|
||
| @router.post("/models/{model}:generateContent") | ||
| async def chat(model: str, request:ChatRequest = Body()) -> dict: |
There was a problem hiding this comment.
MyPy 类型错误: Missing return statement (return)
详细信息请参考 mypy 文档。
feature: 进一步完善了adapter的基本测试用例。
|
lgtm |
2 participants
feat: 未各大模型补充了embed方法
feat: 为LLMBackendAdapter添加了embed抽象方法。
feat: 增加了一个纯embedding模型服务商
tips: 新的服务商暂未注册，等待llm注册逻辑更新。
好的，这是翻译成中文的 pull request 总结：
Sourcery 总结
为多个 LLM 后端适配器添加 Embedding 支持，引入跨不同 AI 提供商的标准化 Embedding 请求和响应格式。
新特性：
增强功能：
Original summary in English
好的，这是翻译成中文的 pull request 总结：
Sourcery 总结
为 LLM 适配器添加嵌入 (embedding) 和重排序 (reranking) 功能。
新功能：
增强：
LLMEmbeddingProtocol和LLMReRankProtocol。LLMEmbeddingRequest/Response,LLMReRankRequest/Response)。构建：
setup.py中删除未使用的依赖项。测试：
Original summary in English
Summary by Sourcery
Add embedding and reranking capabilities to LLM adapters.
New Features:
Enhancements:
LLMEmbeddingProtocolandLLMReRankProtocolfor adapters.LLMEmbeddingRequest/Response,LLMReRankRequest/Response).Build:
setup.py.Tests: