Diuse/zhenxun_bot
1
0
Fork 0
mirror of https://github.com/zhenxun-org/zhenxun_bot.git synced 2026-10-05 20:00:04 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

♻️ refactor(llm): 重构 LLM 服务架构，引入中间件与组件化适配器

Browse Source 
- 【重构】LLM 服务核心架构：
    - 引入中间件管道，统一处理请求生命周期（重试、密钥选择、日志、网络请求）。
    - 适配器重构为组件化设计，分离配置映射、消息转换、响应解析和工具序列化逻辑。
    - 移除 `with_smart_retry` 装饰器，其功能由中间件接管。
    - 移除 `LLMToolExecutor`，工具执行逻辑集成到 `ToolInvoker`。
- 【功能】增强配置系统：
    - `LLMGenerationConfig` 采用组件化结构（Core, Reasoning, Visual, Output, Safety, ToolConfig）。
    - 新增 `GenConfigBuilder` 提供语义化配置构建方式。
    - 新增 `LLMEmbeddingConfig` 用于嵌入专用配置。
    - `CommonOverrides` 迁移并更新至新配置结构。
- 【功能】强化工具系统：
    - 引入 `ToolInvoker` 实现更灵活的工具执行，支持回调与结构化错误。
    - `function_tool` 装饰器支持动态 Pydantic 模型创建和依赖注入 (`ToolParam`, `RunContext`)。
    - 平台原生工具支持 (`GeminiCodeExecution`, `GeminiGoogleSearch`, `GeminiUrlContext`)。
- 【功能】高级生成与嵌入：
    - `generate_structured` 方法支持 In-Context Validation and Repair (IVR) 循环和 AutoCoT (思维链) 包装。
    - 新增 `embed_query` 和 `embed_documents` 便捷嵌入 API。
    - `OpenAIImageAdapter` 支持 OpenAI 兼容的图像生成。
    - `SmartAdapter` 实现模型名称智能路由。
- 【重构】消息与类型系统：
    - `LLMContentPart` 扩展支持更多模态和代码执行相关内容。
    - `LLMMessage` 和 `LLMResponse` 结构更新，支持 `content_parts` 和思维链签名。
    - 统一 `LLMErrorCode` 和用户友好错误消息，提供更详细的网络/代理错误提示。
    - `pyproject.toml` 移除 `bilireq`，新增 `json_repair`。
- 【优化】日志与调试：
    - 引入 `DebugLogOptions`，提供细粒度日志脱敏控制。
    - 增强日志净化器，处理更多敏感数据和长字符串。
- 【清理】删除废弃模块：
    - `zhenxun/services/llm/memory.py`
    - `zhenxun/services/llm/executor.py`
    - `zhenxun/services/llm/config/presets.py`
    - `zhenxun/services/llm/types/content.py`
    - `zhenxun/services/llm/types/enums.py`
    - `zhenxun/services/llm/tools/__init__.py`
    - `zhenxun/services/llm/tools/manager.py`
This commit is contained in:
webjoin111
2025-12-07 18:57:55 +08:00
parent e5b2a872d3
commit bba90e62db
35 changed files with 6087 additions and 3097 deletions
Download Patch File Download Diff File Expand all files Collapse all files
zhenxun/utils/log_sanitizer.py
+164 -19
View File
@@ -14,9 +14,34 @@ def _truncate_base64_string(value: str, threshold: int = 256) -> str:
if value.startswith(prefixes) and len(value) > threshold:
prefix = next((p for p in prefixes if value.startswith(p)), "base64")
return f"[{prefix}_data_omitted_len={len(value)}]"
if len(value) > 1000:
return f"[long_string_omitted_len={len(value)}] {value[:20]}...{value[-20:]}"
if len(value) > 2000:
return f"[long_string_omitted_len={len(value)}] {value[:50]}...{value[-20:]}"
return value
def _truncate_vector_list(vector: list, threshold: int = 10) -> list:
"""如果列表过长（通常是embedding向量），则截断它用于日志显示。"""
if isinstance(vector, list) and len(vector) > threshold:
return [*vector[:3], f"...({len(vector)} floats omitted)...", *vector[-3:]]
return vector
def _recursive_sanitize_any(obj: Any) -> Any:
"""递归清洗任何对象中的长字符串"""
if isinstance(obj, dict):
return {k: _recursive_sanitize_any(v) for k, v in obj.items()}
elif isinstance(obj, list):
return [_recursive_sanitize_any(v) for v in obj]
elif isinstance(obj, str):
return _truncate_base64_string(obj)
return obj
def _sanitize_ui_html(html_string: str) -> str:
"""
专门用于净化UI渲染调试HTML的函数。
@@ -64,6 +89,37 @@ def _sanitize_openai_response(response_json: dict) -> dict:
message["images"][i]["image_url"]["url"] = (
_truncate_base64_string(url)
)
if "reasoning_details" in message and isinstance(
message["reasoning_details"], list
):
for detail in message["reasoning_details"]:
if isinstance(detail, dict):
if "data" in detail and isinstance(detail["data"], str):
if len(detail["data"]) > 100:
detail["data"] = (
f"[encrypted_data_omitted_len={len(detail['data'])}]"
)
if "text" in detail and isinstance(detail["text"], str):
detail["text"] = _truncate_base64_string(
detail["text"], threshold=2000
)
if "data" in sanitized_json and isinstance(sanitized_json["data"], list):
for item in sanitized_json["data"]:
if "embedding" in item and isinstance(item["embedding"], list):
item["embedding"] = _truncate_vector_list(item["embedding"])
if "b64_json" in item and isinstance(item["b64_json"], str):
if len(item["b64_json"]) > 256:
item["b64_json"] = (
f"[base64_json_omitted_len={len(item['b64_json'])}]"
)
if "input" in sanitized_json and isinstance(sanitized_json["input"], list):
for item in sanitized_json["input"]:
if "content" in item and isinstance(item["content"], list):
for part in item["content"]:
if isinstance(part, dict) and part.get("type") == "input_image":
image_url = part.get("image_url")
if isinstance(image_url, str):
part["image_url"] = _truncate_base64_string(image_url)
return sanitized_json
except Exception:
return response_json
@@ -71,22 +127,44 @@ def _sanitize_openai_response(response_json: dict) -> dict:
def _sanitize_openai_request(body: dict) -> dict:
"""净化OpenAI兼容API的请求体，主要截断图片base64。"""
from zhenxun.services.llm.config.providers import (
DebugLogOptions,
get_llm_config,
)
debug_conf = get_llm_config().debug_log
if isinstance(debug_conf, bool):
debug_conf = DebugLogOptions(
show_tools=debug_conf, show_schema=debug_conf, show_safety=debug_conf
)
try:
sanitized_json = copy.deepcopy(body)
if "messages" in sanitized_json and isinstance(
sanitized_json["messages"], list
):
for message in sanitized_json["messages"]:
if "content" in message and isinstance(message["content"], list):
for i, part in enumerate(message["content"]):
if part.get("type") == "image_url":
if "image_url" in part and isinstance(
part["image_url"], dict
):
url = part["image_url"].get("url", "")
message["content"][i]["image_url"]["url"] = (
_truncate_base64_string(url)
)
sanitized_json = _recursive_sanitize_any(copy.deepcopy(body))
if "tools" in sanitized_json and not debug_conf.show_tools:
tools = sanitized_json["tools"]
if isinstance(tools, list):
tool_names = []
for t in tools:
if isinstance(t, dict):
name = None
if "function" in t and isinstance(t["function"], dict):
name = t["function"].get("name")
if not name and "name" in t:
name = t.get("name")
tool_names.append(name or "unknown")
sanitized_json["tools"] = (
f"<{len(tool_names)} tools hidden: {', '.join(tool_names)}>"
)
if "response_format" in sanitized_json and not debug_conf.show_schema:
response_format = sanitized_json["response_format"]
if isinstance(response_format, dict):
if response_format.get("type") == "json_schema":
sanitized_json["response_format"] = {
"type": "json_schema",
"json_schema": "<JSON Schema Hidden>",
}
return sanitized_json
except Exception:
return body
@@ -94,6 +172,9 @@ def _sanitize_openai_request(body: dict) -> dict:
def _sanitize_gemini_response(response_json: dict) -> dict:
"""净化Gemini API的响应体，处理文本和图片生成两种格式。"""
from zhenxun.services.llm.config.providers import get_llm_config
debug_mode = get_llm_config().debug_log
try:
sanitized_json = copy.deepcopy(response_json)
@@ -114,6 +195,15 @@ def _sanitize_gemini_response(response_json: dict) -> dict:
content["parts"][i]["inlineData"]["data"] = (
f"[base64_data_omitted_len={len(data)}]"
)
if "thoughtSignature" in part:
signature = part.get("thoughtSignature", "")
if isinstance(signature, str) and len(signature) > 256:
content["parts"][i]["thoughtSignature"] = (
f"[signature_omitted_len={len(signature)}]"
)
if not debug_mode and isinstance(candidate, dict):
if "safetyRatings" in candidate:
candidate["safetyRatings"] = "<Safety Ratings Hidden>"
if "candidates" in sanitized_json:
_process_candidates(sanitized_json["candidates"])
@@ -124,6 +214,19 @@ def _sanitize_gemini_response(response_json: dict) -> dict:
if "candidates" in sanitized_json["image_generation"]:
_process_candidates(sanitized_json["image_generation"]["candidates"])
if "embeddings" in sanitized_json and isinstance(
sanitized_json["embeddings"], list
):
for embedding in sanitized_json["embeddings"]:
if "values" in embedding and isinstance(embedding["values"], list):
embedding["values"] = _truncate_vector_list(embedding["values"])
if not debug_mode and "promptFeedback" in sanitized_json:
prompt_feedback = sanitized_json.get("promptFeedback") or {}
if isinstance(prompt_feedback, dict) and "safetyRatings" in prompt_feedback:
prompt_feedback["safetyRatings"] = "<Safety Ratings Hidden>"
sanitized_json["promptFeedback"] = prompt_feedback
return sanitized_json
except Exception:
return response_json
@@ -131,8 +234,46 @@ def _sanitize_gemini_response(response_json: dict) -> dict:
def _sanitize_gemini_request(body: dict) -> dict:
"""净化Gemini API的请求体，进行结构转换和总结。"""
from zhenxun.services.llm.config.providers import (
DebugLogOptions,
get_llm_config,
)
debug_conf = get_llm_config().debug_log
if isinstance(debug_conf, bool):
debug_conf = DebugLogOptions(
show_tools=debug_conf, show_schema=debug_conf, show_safety=debug_conf
)
try:
sanitized_body = copy.deepcopy(body)
if "tools" in sanitized_body and not debug_conf.show_tools:
tool_summary = []
for tool_group in sanitized_body["tools"]:
if (
isinstance(tool_group, dict)
and "functionDeclarations" in tool_group
):
declarations = tool_group["functionDeclarations"]
if isinstance(declarations, list):
for func in declarations:
if isinstance(func, dict):
tool_summary.append(func.get("name", "unknown"))
sanitized_body["tools"] = (
f"<{len(tool_summary)} functions hidden: {', '.join(tool_summary)}>"
)
if not debug_conf.show_safety and "safetySettings" in sanitized_body:
sanitized_body["safetySettings"] = "<Safety Settings Hidden>"
if not debug_conf.show_schema and "generationConfig" in sanitized_body:
generation_config = sanitized_body["generationConfig"]
if (
isinstance(generation_config, dict)
and "responseJsonSchema" in generation_config
):
generation_config["responseJsonSchema"] = "<JSON Schema Hidden>"
if "contents" in sanitized_body and isinstance(
sanitized_body["contents"], list
):
@@ -153,6 +294,13 @@ def _sanitize_gemini_request(body: dict) -> dict:
continue
new_parts.append(part)
if "thoughtSignature" in part:
sig = part["thoughtSignature"]
if isinstance(sig, str) and len(sig) > 64:
part["thoughtSignature"] = (
f"[signature_omitted_len={len(sig)}]"
)
if media_summary:
summary_text = (
f"[多模态内容: {len(media_summary)}个文件 - "
@@ -195,8 +343,5 @@ def sanitize_for_logging(data: Any, context: str | None = None) -> Any:
elif context == "ui_html":
if isinstance(data, str):
return _sanitize_ui_html(data)
else:
if isinstance(data, str):
return _truncate_base64_string(data)
return data
return _recursive_sanitize_any(data)
zhenxun/utils/pydantic_compat.py
+24 -20
View File
@@ -10,8 +10,14 @@ from enum import Enum
from pathlib import Path
from typing import Any, TypeVar, get_args, get_origin
from nonebot.compat import PYDANTIC_V2, model_dump
from pydantic import VERSION, BaseModel
from nonebot.compat import (
PYDANTIC_V2,
model_dump,
model_fields,
type_validate_json,
type_validate_python,
)
from pydantic import BaseModel
import ujson as json
T = TypeVar("T", bound=BaseModel)
@@ -27,9 +33,13 @@ __all__ = [
"model_construct",
"model_copy",
"model_dump",
"model_dump_json",
"model_fields",
"model_json_schema",
"model_validate",
"parse_as",
"type_validate_json",
"type_validate_python",
]
@@ -58,12 +68,18 @@ def model_construct(model_class: type[T], **kwargs: Any) -> T:
def model_validate(model_class: type[T], obj: Any) -> T:
"""
Pydantic `model_validate` (v2) 与 `parse_obj` (v1) 的兼容函数。
Pydantic 模型验证兼容函数。
"""
return type_validate_python(model_class, obj)
def model_dump_json(model: BaseModel, **kwargs: Any) -> str:
"""
Pydantic `model.json()` (v1) 和 `model.model_dump_json()` (v2) 的兼容函数。
"""
if PYDANTIC_V2:
return model_class.model_validate(obj)
else:
return model_class.parse_obj(obj)
return model.model_dump_json(**kwargs)
return model.json(**kwargs)
if PYDANTIC_V2:
@@ -78,8 +94,7 @@ def model_json_schema(model_class: type[BaseModel], **kwargs: Any) -> dict[str,
"""
if PYDANTIC_V2:
return model_class.model_json_schema(**kwargs)
else:
return model_class.schema(by_alias=kwargs.get("by_alias", True))
return model_class.schema(by_alias=kwargs.get("by_alias", True))
def _is_pydantic_type(t: Any) -> bool:
@@ -108,18 +123,7 @@ def _dump_pydantic_obj(obj: Any) -> Any:
return obj
def parse_as(type_: type[V], obj: Any) -> V:
"""
一个兼容 Pydantic V1 的 parse_obj_as 和V2的TypeAdapter.validate_python 的辅助函数。
"""
if VERSION.startswith("1"):
from pydantic import parse_obj_as
return parse_obj_as(type_, obj)
else:
from pydantic import TypeAdapter # type: ignore
return TypeAdapter(type_).validate_python(obj)
parse_as = type_validate_python
def dump_json_safely(obj: Any, **kwargs) -> str: