fix: 修复Gemini 2.5思考模型工具调用时thought_signature缺失导致400错误

- Google provider统一使用ChatGoogleGenerativeAI原生接口,不再走OpenAI兼容端点
  (OpenAI协议不支持thought_signature字段,导致思考模型工具调用必然失败)
- 通过client_args传递代理配置,替代原来的OpenAI兼容端点+openai_proxy方案
- 修补langchain-google-genai的_is_gemini_3_or_later()以覆盖Gemini 2.5模型
- 自动适配httpx代理参数名(proxies/proxy),修复代理配置被静默丢弃的问题
This commit is contained in:
jxxghp
2026-04-04 07:24:47 +08:00
parent 974d5f2f49
commit 5b892b3a63

View File

@@ -1,11 +1,61 @@
"""LLM模型相关辅助功能"""
import inspect
from typing import List
from app.core.config import settings
from app.log import logger
def _patch_gemini_thought_signature():
"""
修复 langchain-google-genai 中 Gemini 2.5 思考模型的 thought_signature 兼容问题。
langchain-google-genai 的 _is_gemini_3_or_later() 仅检查 "gemini-3"
导致 Gemini 2.5 思考模型(如 gemini-2.5-flash、gemini-2.5-pro在工具调用时
缺少 thought_signature 而报错 400。
此补丁将检查范围扩展到 Gemini 2.5 模型。
"""
try:
import langchain_google_genai.chat_models as _cm
# 仅在未修补时执行
if getattr(_cm, "_thought_signature_patched", False):
return
def _patched_is_gemini_3_or_later(model_name: str) -> bool:
if not model_name:
return False
name = model_name.lower().replace("models/", "")
# Gemini 2.5 思考模型也需要 thought_signature 支持
return "gemini-3" in name or "gemini-2.5" in name
_cm._is_gemini_3_or_later = _patched_is_gemini_3_or_later
_cm._thought_signature_patched = True
logger.debug(
"已修补 langchain-google-genai thought_signature 兼容性(覆盖 Gemini 2.5 模型)"
)
except Exception as e:
logger.warning(f"修补 langchain-google-genai thought_signature 失败: {e}")
def _get_httpx_proxy_key() -> str:
"""
获取当前 httpx 版本支持的代理参数名。
httpx < 0.28 使用 "proxies"(复数),>= 0.28 使用 "proxy"(单数)。
google-genai SDK 会静默过滤掉不在 httpx.Client.__init__ 签名中的参数,
因此必须使用与当前 httpx 版本匹配的参数名。
"""
try:
import httpx
params = inspect.signature(httpx.Client.__init__).parameters
if "proxy" in params:
return "proxy"
return "proxies"
except Exception:
return "proxies"
class LLMHelper:
"""LLM模型相关辅助功能"""
@@ -23,31 +73,27 @@ class LLMHelper:
raise ValueError("未配置LLM API Key")
if provider == "google":
# 修补 Gemini 2.5 思考模型的 thought_signature 兼容性
_patch_gemini_thought_signature()
# 统一使用 langchain-google-genai 原生接口
# 不使用 OpenAI 兼容端点,因其不支持 Gemini 思考模型的 thought_signature
# 会导致工具调用时报错 400
from langchain_google_genai import ChatGoogleGenerativeAI
client_args = None
if settings.PROXY_HOST:
# 通过代理使用 Google 的 OpenAI 兼容接口
from langchain_openai import ChatOpenAI
proxy_key = _get_httpx_proxy_key()
client_args = {proxy_key: settings.PROXY_HOST}
model = ChatOpenAI(
model=settings.LLM_MODEL,
api_key=api_key,
max_retries=3,
base_url="https://generativelanguage.googleapis.com/v1beta/openai",
temperature=settings.LLM_TEMPERATURE,
streaming=streaming,
stream_usage=True,
openai_proxy=settings.PROXY_HOST,
)
else:
# 使用 langchain-google-genai 原生接口v4 API 变更google_api_key → api_keymax_retries → retries
from langchain_google_genai import ChatGoogleGenerativeAI
model = ChatGoogleGenerativeAI(
model=settings.LLM_MODEL,
api_key=api_key,
retries=3,
temperature=settings.LLM_TEMPERATURE,
streaming=streaming,
)
model = ChatGoogleGenerativeAI(
model=settings.LLM_MODEL,
api_key=api_key,
retries=3,
temperature=settings.LLM_TEMPERATURE,
streaming=streaming,
client_args=client_args,
)
elif provider == "deepseek":
from langchain_deepseek import ChatDeepSeek
@@ -103,9 +149,11 @@ class LLMHelper:
http_options = None
if settings.PROXY_HOST:
proxy_key = _get_httpx_proxy_key()
proxy_args = {proxy_key: settings.PROXY_HOST}
http_options = HttpOptions(
client_args={"proxy": settings.PROXY_HOST},
async_client_args={"proxy": settings.PROXY_HOST},
client_args=proxy_args,
async_client_args=proxy_args,
)
client = genai.Client(api_key=api_key, http_options=http_options)