mirror of
https://github.com/zhenxun-org/zhenxun_bot.git
synced 2026-10-18 02:30:08 +08:00
✨ feat(llm): 增强LLM服务，支持图片生成、响应验证与OpenRouter集成 (#2054)
* ✨ feat(llm): 增强LLM服务，支持图片生成、响应验证与OpenRouter集成 - 【新功能】统一图片生成与编辑API `create_image`，支持文生图、图生图及多图输入 - 【新功能】引入LLM响应验证机制，通过 `validation_policy` 和 `response_validator` 确保响应内容符合预期，例如强制返回图片 - 【新功能】适配OpenRouter API，扩展LLM服务提供商支持，并添加OpenRouter特定请求头 - 【重构】将日志净化逻辑重构至 `log_sanitizer` 模块，提供统一的净化入口，并应用于NoneBot消息、LLM请求/响应日志 - 【修复】优化Gemini适配器，正确解析图片生成响应中的Base64图片数据，并更新模型能力注册表 * ✨ feat(image): 优化图片生成响应并返回完整LLMResponse * ✨ feat(llm): 为 OpenAI 兼容请求体添加日志净化 * 🐛 fix(ui): 截断UI调试HTML日志中的长base64图片数据 --------- Co-authored-by: webjoin111 <455457521@qq.com>
This commit is contained in:
Rumio
@@ -12,6 +12,7 @@ from typing import Any, TypeVar
|
||||
from pydantic import BaseModel
|
||||
|
||||
from zhenxun.services.log import logger
|
||||
from zhenxun.utils.log_sanitizer import sanitize_for_logging
|
||||
|
||||
from .adapters.base import RequestData
|
||||
from .config import LLMGenerationConfig
|
||||
@@ -34,7 +35,6 @@ from .types import (
|
||||
ToolExecutable,
|
||||
)
|
||||
from .types.capabilities import ModelCapabilities, ModelModality
|
||||
from .utils import _sanitize_request_body_for_logging
|
||||
|
||||
T = TypeVar("T", bound=BaseModel)
|
||||
|
||||
@@ -187,7 +187,13 @@ class LLMModel(LLMModelBase):
|
||||
logger.debug(f"🔑 API密钥: {masked_key}")
|
||||
logger.debug(f"📋 请求头: {dict(request_data.headers)}")
|
||||
|
||||
sanitized_body = _sanitize_request_body_for_logging(request_data.body)
|
||||
sanitizer_req_context_map = {"gemini": "gemini_request"}
|
||||
sanitizer_req_context = sanitizer_req_context_map.get(
|
||||
self.api_type, "openai_request"
|
||||
)
|
||||
sanitized_body = sanitize_for_logging(
|
||||
request_data.body, context=sanitizer_req_context
|
||||
)
|
||||
request_body_str = json.dumps(sanitized_body, ensure_ascii=False, indent=2)
|
||||
logger.debug(f"📦 请求体: {request_body_str}")
|
||||
|
||||
@@ -200,8 +206,11 @@ class LLMModel(LLMModelBase):
|
||||
logger.debug(f"📥 响应状态码: {http_response.status_code}")
|
||||
logger.debug(f"📄 响应头: {dict(http_response.headers)}")
|
||||
|
||||
response_bytes = await http_response.aread()
|
||||
logger.debug(f"📦 响应体已完整读取 ({len(response_bytes)} bytes)")
|
||||
|
||||
if http_response.status_code != 200:
|
||||
error_text = http_response.text
|
||||
error_text = response_bytes.decode("utf-8", errors="ignore")
|
||||
logger.error(
|
||||
f"❌ HTTP请求失败: {http_response.status_code} - {error_text} "
|
||||
f"[{log_context}]"
|
||||
@@ -232,13 +241,22 @@ class LLMModel(LLMModelBase):
|
||||
)
|
||||
|
||||
try:
|
||||
response_json = http_response.json()
|
||||
response_json = json.loads(response_bytes)
|
||||
|
||||
sanitizer_context_map = {"gemini": "gemini_response"}
|
||||
sanitizer_context = sanitizer_context_map.get(
|
||||
self.api_type, "openai_response"
|
||||
)
|
||||
|
||||
sanitized_for_log = sanitize_for_logging(
|
||||
response_json, context=sanitizer_context
|
||||
)
|
||||
|
||||
response_json_str = json.dumps(
|
||||
response_json, ensure_ascii=False, indent=2
|
||||
sanitized_for_log, ensure_ascii=False, indent=2
|
||||
)
|
||||
logger.debug(f"📋 响应JSON: {response_json_str}")
|
||||
parsed_data = parse_response_func(response_json)
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"解析 {log_context} 响应失败: {e}", e=e)
|
||||
await self.key_store.record_failure(api_key, None, str(e))
|
||||
@@ -290,7 +308,7 @@ class LLMModel(LLMModelBase):
|
||||
adapter.validate_embedding_response(response_json)
|
||||
return adapter.parse_embedding_response(response_json)
|
||||
|
||||
parsed_data, api_key_used = await self._perform_api_call(
|
||||
parsed_data, _api_key_used = await self._perform_api_call(
|
||||
prepare_request_func=prepare_request,
|
||||
parse_response_func=parse_response,
|
||||
http_client=http_client,
|
||||
@@ -376,6 +394,7 @@ class LLMModel(LLMModelBase):
|
||||
return LLMResponse(
|
||||
text=response_data.text,
|
||||
usage_info=response_data.usage_info,
|
||||
image_bytes=response_data.image_bytes,
|
||||
raw_response=response_data.raw_response,
|
||||
tool_calls=response_tool_calls if response_tool_calls else None,
|
||||
code_executions=response_data.code_executions,
|
||||
@@ -390,6 +409,56 @@ class LLMModel(LLMModelBase):
|
||||
failed_keys=failed_keys,
|
||||
log_context="Generation",
|
||||
)
|
||||
|
||||
if config:
|
||||
if config.response_validator:
|
||||
try:
|
||||
config.response_validator(parsed_data)
|
||||
except Exception as e:
|
||||
raise LLMException(
|
||||
f"响应内容未通过自定义验证器: {e}",
|
||||
code=LLMErrorCode.API_RESPONSE_INVALID,
|
||||
details={"validator_error": str(e)},
|
||||
cause=e,
|
||||
) from e
|
||||
|
||||
policy = config.validation_policy
|
||||
if policy:
|
||||
if policy.get("require_image") and not parsed_data.image_bytes:
|
||||
if self.api_type == "gemini" and parsed_data.raw_response:
|
||||
usage_metadata = parsed_data.raw_response.get(
|
||||
"usageMetadata", {}
|
||||
)
|
||||
prompt_token_details = usage_metadata.get(
|
||||
"promptTokensDetails", []
|
||||
)
|
||||
prompt_had_image = any(
|
||||
detail.get("modality") == "IMAGE"
|
||||
for detail in prompt_token_details
|
||||
)
|
||||
|
||||
if prompt_had_image:
|
||||
raise LLMException(
|
||||
"响应验证失败：模型接收了图片输入但未生成图片。",
|
||||
code=LLMErrorCode.API_RESPONSE_INVALID,
|
||||
details={
|
||||
"policy": policy,
|
||||
"text_response": parsed_data.text,
|
||||
"raw_response": parsed_data.raw_response,
|
||||
},
|
||||
)
|
||||
else:
|
||||
logger.debug("Gemini提示词中未包含图片，跳过图片要求重试。")
|
||||
else:
|
||||
raise LLMException(
|
||||
"响应验证失败：要求返回图片但未找到图片数据。",
|
||||
code=LLMErrorCode.API_RESPONSE_INVALID,
|
||||
details={
|
||||
"policy": policy,
|
||||
"text_response": parsed_data.text,
|
||||
},
|
||||
)
|
||||
|
||||
return parsed_data, api_key_used
|
||||
|
||||
async def close(self):
|
||||
|
||||
Reference in New Issue
Block a user