From 5b892b3a63878c207c0fd1e733f0ed99b2c230d1 Mon Sep 17 00:00:00 2001
From: jxxghp <jxxghp@gmail.com>
Date: Sat, 4 Apr 2026 07:24:47 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20=E4=BF=AE=E5=A4=8DGemini=202.5=E6=80=9D?=
 =?UTF-8?q?=E8=80=83=E6=A8=A1=E5=9E=8B=E5=B7=A5=E5=85=B7=E8=B0=83=E7=94=A8?=
 =?UTF-8?q?=E6=97=B6thought=5Fsignature=E7=BC=BA=E5=A4=B1=E5=AF=BC?=
 =?UTF-8?q?=E8=87=B4400=E9=94=99=E8=AF=AF?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- 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
---
 app/helper/llm.py | 98 +++++++++++++++++++++++++++++++++++------------
 1 file changed, 73 insertions(+), 25 deletions(-)

diff --git a/app/helper/llm.py b/app/helper/llm.py
index 5526d020..2c570499 100644
--- a/app/helper/llm.py
+++ b/app/helper/llm.py
@@ -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_key，max_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)