75 lines
2.8 KiB
Python
75 lines
2.8 KiB
Python
|
"""从客服对话中提取长期记忆候选。"""
|
||
|
|
|
||
|
|
from __future__ import annotations
|
||
|
|
|
||
|
|
import json
|
||
|
|
import re
|
||
|
|
from typing import Any
|
||
|
|
|
||
|
|
from service.memory.schemas import MemorySource, MemoryType
|
||
|
|
|
||
|
|
|
||
|
|
class DialogueMemoryExtractor:
|
||
|
|
"""使用项目统一 LLM 提取结构化客户记忆候选。"""
|
||
|
|
|
||
|
|
SYSTEM_PROMPT = """
|
||
|
|
你是客服记忆提取器，只提取用户明确表达或稳定陈述的客户信息。
|
||
|
|
只返回 JSON 数组，不要输出 Markdown 或解释文字。
|
||
|
|
每项必须包含：tag、content、memory_type、source。
|
||
|
|
memory_type 只能是 PROFILE_FACT、PROFILE_CANDIDATE、CUSTOMER_PREFERENCE、INVESTMENT_GOAL、SERVICE_FACT。
|
||
|
|
source 只能是 dialogue_confirmed、dialogue_stated、dialogue_inferred。
|
||
|
|
客服知识问题、产品政策、寒暄、客服回复内容不要提取。
|
||
|
|
不确定的信息使用 dialogue_inferred，无法形成客户画像的信息不要提取。
|
||
|
|
""".strip()
|
||
|
|
|
||
|
|
def __init__(self, llm_client):
|
||
|
|
self.llm_client = llm_client
|
||
|
|
|
||
|
|
async def extract(self, query: str, *, context: dict[str, Any] | None = None) -> list[dict[str, str]]:
|
||
|
|
"""提取并校验本轮用户消息中的客户记忆候选。"""
|
||
|
|
prompt = [{"role": "system", "content": self.SYSTEM_PROMPT}]
|
||
|
|
prompt.append(
|
||
|
|
{
|
||
|
|
"role": "user",
|
||
|
|
"content": json.dumps(
|
||
|
|
{"query": query, "existing_memory": context or {}},
|
||
|
|
ensure_ascii=False,
|
||
|
|
),
|
||
|
|
}
|
||
|
|
)
|
||
|
|
response = await self.llm_client.chat(prompt, temperature=0, max_tokens=800)
|
||
|
|
return self._parse(response)
|
||
|
|
|
||
|
|
@staticmethod
|
||
|
|
def _parse(response: str) -> list[dict[str, str]]:
|
||
|
|
"""解析 LLM JSON，并过滤不符合记忆契约的内容。"""
|
||
|
|
text = response.strip()
|
||
|
|
fenced = re.search(r"```(?:json)?\s*(.*?)\s*```", text, re.S | re.I)
|
||
|
|
if fenced:
|
||
|
|
text = fenced.group(1)
|
||
|
|
data = json.loads(text)
|
||
|
|
if not isinstance(data, list):
|
||
|
|
raise ValueError("记忆提取结果必须是数组")
|
||
|
|
valid_types = {item.value for item in MemoryType}
|
||
|
|
valid_sources = {item.value for item in MemorySource}
|
||
|
|
result = []
|
||
|
|
for item in data:
|
||
|
|
if not isinstance(item, dict):
|
||
|
|
continue
|
||
|
|
if not all(item.get(key) for key in ("tag", "content", "memory_type", "source")):
|
||
|
|
continue
|
||
|
|
if item["memory_type"] not in valid_types or item["source"] not in valid_sources:
|
||
|
|
continue
|
||
|
|
result.append(
|
||
|
|
{
|
||
|
|
"tag": str(item["tag"])[:64],
|
||
|
|
"content": str(item["content"])[:512],
|
||
|
|
"memory_type": item["memory_type"],
|
||
|
|
"source": item["source"],
|
||
|
|
}
|
||
|
|
)
|
||
|
|
return result
|
||
|
|
|
||
|
|
|
||
|
|
__all__ = ["DialogueMemoryExtractor"]
|