243 lines
8.7 KiB
Python
243 lines
8.7 KiB
Python
"""
|
|
思维链控制模块 - 重构版
|
|
统一管理各模型的思维链/推理模式关闭参数
|
|
|
|
设计原则:
|
|
- 只认模型特征,不认服务商(Ollama除外,因其使用原生API)
|
|
- 按参数类型分组,而非按服务商分组
|
|
- 支持主流平台:智谱/Qwen/DeepSeek/Gemini/Grok/Ollama
|
|
- 保留400错误自动降级机制(在openai_base.py中)
|
|
"""
|
|
|
|
import re
|
|
from typing import Dict, Any, List
|
|
|
|
|
|
# ==================== 一级:通用模型规则(与服务商无关) ====================
|
|
# 按参数类型分组,遍历匹配
|
|
|
|
THINKING_CONTROL_RULES: List[Dict[str, Any]] = [
|
|
{
|
|
"name": "zhipu_glm",
|
|
"description": "智谱GLM-4.5/4.6系列",
|
|
"patterns": [r"glm[-_/.]?4\.(5|6)"],
|
|
"params": {"thinking": {"type": "disabled"}},
|
|
"sources": ["智谱官方API", "各聚合平台"]
|
|
},
|
|
{
|
|
"name": "qwen3_enable_thinking",
|
|
"description": "Qwen3系列(支持enable_thinking平台)",
|
|
"patterns": [
|
|
r"qwen[-_/.]?3(?!.*thinking)(?!.*r1)", # Qwen3非thinking/r1变体
|
|
r"qwen.*[-_/.]?vl", # Qwen-VL系列(包括qwen2/qwen3-vl)
|
|
],
|
|
"params": {"enable_thinking": False},
|
|
"sources": ["魔搭", "阿里百炼", "硅基流动"]
|
|
},
|
|
{
|
|
"name": "qwen_deepseek_reasoning",
|
|
"description": "Qwen/DeepSeek推理模型(R1系列)",
|
|
"patterns": [
|
|
r"qwen.*[-_/.]?(r1|thinking)", # Qwen推理模型
|
|
r"deepseek.*[-_/.]?(r1|reason)", # DeepSeek R1系列
|
|
],
|
|
"params": {"reasoning": {"effort": "none"}},
|
|
"sources": ["OpenRouter", "302.ai", "其他聚合平台"]
|
|
},
|
|
{
|
|
"name": "deepseek_v3_thinking",
|
|
"description": "DeepSeek V3/V3.2官方API",
|
|
"patterns": [
|
|
r"deepseek[-_/.]?v3(\.1|\.2)?(?!.*r1)", # V3系列非R1
|
|
r"deepseek[-_/.]?chat", # DeepSeek Chat
|
|
],
|
|
"params": {"thinking": {"type": "disabled"}},
|
|
"sources": ["DeepSeek官方API"],
|
|
"notes": "也可通过模型名加/thinking控制"
|
|
},
|
|
{
|
|
"name": "gemini2_reasoning",
|
|
"description": "Gemini 2.0/2.5 Flash/Lite",
|
|
"patterns": [r"gemini[-_/.]?2\.(0|5)[-_/.]?(flash|lite)"],
|
|
"params": {"reasoning_effort": "none"},
|
|
"sources": ["Google官方API", "OpenAI兼容接口"]
|
|
},
|
|
{
|
|
"name": "gemini3_thinking",
|
|
"description": "Gemini 3.0系列(通过reasoning_effort控制)",
|
|
"patterns": [r"gemini[-_/.]?3"],
|
|
"params": {"reasoning_effort": "low"},
|
|
"sources": ["Google官方API (OpenAI兼容接口)"],
|
|
"notes": "⚠️ 官方OpenAI端点使用reasoning_effort参数"
|
|
},
|
|
{
|
|
"name": "grok3_mini_reasoning",
|
|
"description": "Grok-3-mini(支持reasoning_effort)",
|
|
"patterns": [r"grok[-_/.]?3[-_/.]?mini"],
|
|
"params": {"reasoning_effort": "low"},
|
|
"sources": ["xAI官方API"]
|
|
},
|
|
# 注意:Grok-4系列是内置推理模型,不支持reasoning_effort参数,已在排除规则中处理
|
|
]
|
|
|
|
|
|
# ==================== 二级:Ollama特殊规则(原生API专用) ====================
|
|
# Ollama使用/api/chat原生接口,参数格式不同
|
|
#
|
|
# ⚠️ 重要发现：think 参数行为与直觉相反
|
|
# - think: true → 关闭思考（不输出 <think> 标签）
|
|
# - think: false → 启用思考（输出 <think> 标签）
|
|
# - 不发送 → 默认关闭思考（Ollama 默认行为）
|
|
|
|
OLLAMA_NATIVE_RULES = {
|
|
"parameter_name": "think",
|
|
"disable_value": True, # disable_thinking=True 时发送此值（关闭思考）
|
|
"enable_value": False, # disable_thinking=False 时发送此值（启用思考）
|
|
"supported_patterns": [
|
|
# r"qwen", # 移除：太宽泛，会误伤qwen-instruct等普通模型
|
|
r"deepseek.*r1", # DeepSeek R1系列
|
|
r"qwen.*r1", # Qwen R1变体
|
|
r".*thinking", # 任何名称包含thinking的模型
|
|
],
|
|
"notes": "Ollama原生API /api/chat 顶层参数"
|
|
}
|
|
|
|
|
|
# ==================== 三级:排除规则(明确不支持关闭的模型) ====================
|
|
# 这些模型要么不支持关闭推理,要么是内置推理模型
|
|
|
|
EXCLUDE_PATTERNS = [
|
|
r"gemini[-_/.]?2\.5[-_/.]?pro", # Gemini 2.5 Pro官方明确不支持关闭
|
|
r"grok[-_/.]?4", # Grok-4系列是内置推理模型,无reasoning_effort参数
|
|
r".*speciale", # DeepSeek Speciale仅thinking模式
|
|
]
|
|
|
|
|
|
# ==================== 核心匹配函数 ====================
|
|
|
|
def build_thinking_suppression(provider: str, model: str, disable_thinking: bool = True) -> Dict[str, Any]:
|
|
"""
|
|
根据模型特征返回思维链控制参数
|
|
|
|
匹配优先级:
|
|
1. 检查排除规则 → 返回空字典(明确不支持的模型)
|
|
2. Ollama特殊处理 → 原生API参数格式
|
|
3. 遍历通用规则 → 返回第一个匹配的参数
|
|
4. 无匹配 → 返回空字典
|
|
|
|
参数:
|
|
provider: 服务商标识(仅用于Ollama判断,其他服务商忽略)
|
|
model: 模型名称
|
|
disable_thinking: True=关闭思考, False=启用思考
|
|
|
|
返回:
|
|
Dict: 思维链控制参数字典,或空字典
|
|
|
|
示例:
|
|
>>> build_thinking_suppression("ollama", "qwen3-thinking", disable_thinking=True)
|
|
{"think": True} # 关闭思考
|
|
|
|
>>> build_thinking_suppression("ollama", "qwen3-thinking", disable_thinking=False)
|
|
{"think": False} # 启用思考
|
|
|
|
>>> build_thinking_suppression("zhipu", "glm-4.5")
|
|
{"thinking": {"type": "disabled"}}
|
|
|
|
>>> build_ thinking_suppression("custom", "gemini-2.5-pro")
|
|
{} # 不支持关闭
|
|
"""
|
|
if not model:
|
|
return {}
|
|
|
|
model_lower = model.strip().lower()
|
|
provider_lower = provider.strip().lower() if provider else ""
|
|
|
|
# 1. 检查排除规则
|
|
for exclude_pattern in EXCLUDE_PATTERNS:
|
|
if re.search(exclude_pattern, model_lower):
|
|
return {} # 明确不支持的模型,返回空字典
|
|
|
|
# 2. Ollama特殊处理(原生API)
|
|
if provider_lower == "ollama":
|
|
for pattern in OLLAMA_NATIVE_RULES["supported_patterns"]:
|
|
if re.search(pattern, model_lower):
|
|
# 根据 disable_thinking 参数选择值
|
|
think_value = OLLAMA_NATIVE_RULES["disable_value"] if disable_thinking else OLLAMA_NATIVE_RULES["enable_value"]
|
|
return {
|
|
OLLAMA_NATIVE_RULES["parameter_name"]: think_value
|
|
}
|
|
return {} # Ollama但不在白名单
|
|
|
|
# 3. 遍历通用规则,返回第一个匹配项
|
|
for rule in THINKING_CONTROL_RULES:
|
|
for pattern in rule["patterns"]:
|
|
if re.search(pattern, model_lower):
|
|
# 返回副本,避免修改原始数据
|
|
return rule["params"].copy()
|
|
|
|
# 4. 无匹配,返回空字典
|
|
return {}
|
|
|
|
|
|
# ==================== 辅助函数(用于调试) ====================
|
|
|
|
def get_rule_info(provider: str, model: str) -> Dict[str, Any]:
|
|
"""
|
|
获取匹配规则的详细信息(用于调试)
|
|
|
|
返回:
|
|
Dict: {
|
|
"matched": bool,
|
|
"rule_name": str,
|
|
"description": str,
|
|
"params": Dict,
|
|
"sources": List[str]
|
|
}
|
|
"""
|
|
if not model:
|
|
return {"matched": False}
|
|
|
|
model_lower = model.strip().lower()
|
|
provider_lower = provider.strip().lower() if provider else ""
|
|
|
|
# 检查排除规则
|
|
for exclude_pattern in EXCLUDE_PATTERNS:
|
|
if re.search(exclude_pattern, model_lower):
|
|
return {
|
|
"matched": True,
|
|
"rule_name": "EXCLUDED",
|
|
"description": "明确不支持关闭思维链的模型",
|
|
"params": {},
|
|
"sources": []
|
|
}
|
|
|
|
# Ollama特殊处理
|
|
if provider_lower == "ollama":
|
|
for pattern in OLLAMA_NATIVE_RULES["supported_patterns"]:
|
|
if re.search(pattern, model_lower):
|
|
return {
|
|
"matched": True,
|
|
"rule_name": "ollama_native",
|
|
"description": "Ollama原生API",
|
|
"params": {
|
|
OLLAMA_NATIVE_RULES["parameter_name"]:
|
|
OLLAMA_NATIVE_RULES["parameter_value"]
|
|
},
|
|
"sources": ["Ollama /api/chat"]
|
|
}
|
|
return {"matched": False}
|
|
|
|
# 通用规则
|
|
for rule in THINKING_CONTROL_RULES:
|
|
for pattern in rule["patterns"]:
|
|
if re.search(pattern, model_lower):
|
|
return {
|
|
"matched": True,
|
|
"rule_name": rule["name"],
|
|
"description": rule["description"],
|
|
"params": rule["params"].copy(),
|
|
"sources": rule.get("sources", [])
|
|
}
|
|
|
|
return {"matched": False}
|