153 lines
4.9 KiB
Python
153 lines
4.9 KiB
Python
"""
|
||
意图识别器
|
||
使用小参数 LLM 进行意图二分类
|
||
"""
|
||
|
||
import os
|
||
import json
|
||
from pathlib import Path
|
||
from typing import Optional
|
||
from dataclasses import dataclass
|
||
from dotenv import load_dotenv
|
||
|
||
from llm.client import get_client, LLMClientError, ENV_PATH
|
||
from llm.prompts import INTENT_CLASSIFICATION_SYSTEM, INTENT_CLASSIFICATION_USER
|
||
from intent.labels import CHAT, EXECUTION, EXECUTION_CONFIDENCE_THRESHOLD, VALID_LABELS
|
||
|
||
|
||
@dataclass
|
||
class IntentResult:
|
||
"""意图识别结果"""
|
||
label: str # chat 或 execution
|
||
confidence: float # 0.0 ~ 1.0
|
||
reason: str # 中文解释
|
||
raw_response: Optional[str] = None # 原始 LLM 响应（调试用）
|
||
|
||
|
||
class IntentClassifier:
|
||
"""
|
||
意图分类器
|
||
|
||
使用小参数 LLM（如 qwen2.5:7b-instruct）进行快速意图识别
|
||
"""
|
||
|
||
def __init__(self):
|
||
load_dotenv(ENV_PATH)
|
||
self.model_name = os.getenv("INTENT_MODEL_NAME")
|
||
|
||
def classify(self, user_input: str) -> IntentResult:
|
||
"""
|
||
对用户输入进行意图分类
|
||
|
||
Args:
|
||
user_input: 用户输入的文本
|
||
|
||
Returns:
|
||
IntentResult: 包含 label, confidence, reason 的结果
|
||
"""
|
||
try:
|
||
client = get_client()
|
||
|
||
messages = [
|
||
{"role": "system", "content": INTENT_CLASSIFICATION_SYSTEM},
|
||
{"role": "user", "content": INTENT_CLASSIFICATION_USER.format(user_input=user_input)}
|
||
]
|
||
|
||
response = client.chat(
|
||
messages=messages,
|
||
model=self.model_name,
|
||
temperature=0.1, # 低温度，更确定性的输出
|
||
max_tokens=256
|
||
)
|
||
|
||
return self._parse_response(response)
|
||
|
||
except LLMClientError as e:
|
||
# LLM 调用失败，走兜底逻辑
|
||
return IntentResult(
|
||
label=CHAT,
|
||
confidence=0.0,
|
||
reason=f"意图识别失败（{str(e)}），默认为对话模式"
|
||
)
|
||
except Exception as e:
|
||
# 其他异常，走兜底逻辑
|
||
return IntentResult(
|
||
label=CHAT,
|
||
confidence=0.0,
|
||
reason=f"意图识别异常（{str(e)}），默认为对话模式"
|
||
)
|
||
|
||
def _parse_response(self, response: str) -> IntentResult:
|
||
"""
|
||
解析 LLM 响应
|
||
|
||
尝试解析 JSON，若失败则走兜底逻辑
|
||
"""
|
||
try:
|
||
# 尝试提取 JSON（LLM 可能会在 JSON 前后加一些文字）
|
||
json_str = self._extract_json(response)
|
||
data = json.loads(json_str)
|
||
|
||
# 验证必要字段
|
||
label = data.get("label", "").lower()
|
||
confidence = float(data.get("confidence", 0.0))
|
||
reason = data.get("reason", "无")
|
||
|
||
# 验证 label 有效性
|
||
if label not in VALID_LABELS:
|
||
return IntentResult(
|
||
label=CHAT,
|
||
confidence=0.0,
|
||
reason=f"无效的意图标签 '{label}'，默认为对话模式",
|
||
raw_response=response
|
||
)
|
||
|
||
# 应用置信度阈值
|
||
if label == EXECUTION and confidence < EXECUTION_CONFIDENCE_THRESHOLD:
|
||
return IntentResult(
|
||
label=CHAT,
|
||
confidence=confidence,
|
||
reason=f"执行任务置信度不足（{confidence:.2f} < {EXECUTION_CONFIDENCE_THRESHOLD}），降级为对话模式。原因: {reason}",
|
||
raw_response=response
|
||
)
|
||
|
||
return IntentResult(
|
||
label=label,
|
||
confidence=confidence,
|
||
reason=reason,
|
||
raw_response=response
|
||
)
|
||
|
||
except (json.JSONDecodeError, ValueError, TypeError) as e:
|
||
# JSON 解析失败，走兜底逻辑
|
||
return IntentResult(
|
||
label=CHAT,
|
||
confidence=0.0,
|
||
reason=f"响应解析失败，默认为对话模式",
|
||
raw_response=response
|
||
)
|
||
|
||
def _extract_json(self, text: str) -> str:
|
||
"""
|
||
从文本中提取 JSON 字符串
|
||
|
||
LLM 可能会在 JSON 前后添加解释文字，需要提取纯 JSON 部分
|
||
"""
|
||
# 尝试找到 JSON 对象的起止位置
|
||
start = text.find('{')
|
||
end = text.rfind('}')
|
||
|
||
if start != -1 and end != -1 and end > start:
|
||
return text[start:end + 1]
|
||
|
||
# 如果找不到，返回原文本让 json.loads 报错
|
||
return text
|
||
|
||
|
||
# 便捷函数
|
||
def classify_intent(user_input: str) -> IntentResult:
|
||
"""快速进行意图分类"""
|
||
classifier = IntentClassifier()
|
||
return classifier.classify(user_input)
|
||
|