109 lines
4.5 KiB
Python
109 lines
4.5 KiB
Python
"""基于当前大屏和澄清历史回答问题"""
|
|
|
|
import json
|
|
|
|
from langchain_core.messages import HumanMessage, SystemMessage
|
|
from langchain_core.output_parsers import StrOutputParser
|
|
from langgraph.config import get_stream_writer
|
|
|
|
from app.clients.llm import get_llm
|
|
from app.processor.base import BaseNode
|
|
from app.processor.state import QueryState
|
|
from app.prompt.qa import OFFER_EXTRACT_PROMPT, QA_SYSTEM_PROMPT
|
|
from app.schemas.offer import ExtractedOffer, InformReply
|
|
|
|
|
|
class QANode(BaseNode):
|
|
name = "qa_node"
|
|
|
|
async def process(self, state: QueryState) -> QueryState:
|
|
# 浅拷贝状态
|
|
updated_state: QueryState = state.copy()
|
|
|
|
try:
|
|
# 图执行上下文中通过 custom 流逐段推送回答，独立调用节点时没有写入器
|
|
writer = get_stream_writer()
|
|
except RuntimeError:
|
|
writer = None
|
|
|
|
def emit(delta: str) -> None:
|
|
if writer is not None:
|
|
writer(
|
|
{
|
|
"type": "answer_delta",
|
|
"message_id": state.get("message_id"),
|
|
"delta": delta,
|
|
}
|
|
)
|
|
|
|
content = state["input_context"].model_dump_json(exclude_unset=True)
|
|
offer = None
|
|
# 请求无法执行或执行失败时，附上事实说明，由本节点如实告知用户
|
|
if notice := state.get("notice"):
|
|
content += f"\n\n系统说明：{notice}"
|
|
# 回复中的替代修改需要结构化保存，下一轮用户同意时直接执行，因此不流式输出
|
|
reply = await get_llm(InformReply).ainvoke(
|
|
[SystemMessage(content=QA_SYSTEM_PROMPT), HumanMessage(content=content)]
|
|
)
|
|
answer, offer = reply.answer.strip(), reply.offer
|
|
emit(answer)
|
|
else:
|
|
answer = await self._stream(content, emit)
|
|
# 如果没有有效文本，抛出异常，避免将空回答标记为处理完成
|
|
if not answer:
|
|
raise ValueError("问答模型未返回文本回答")
|
|
# 回答已流式输出，再提取其中等待用户确认的修改，下一轮同意时直接执行
|
|
offer = await self._extract_offer(state["input_context"].prompt, answer)
|
|
|
|
# 保存最终回答，供后续输出给用户
|
|
updated_state["answer"] = answer
|
|
# 保存等待用户确认的提议，没有提议时清空
|
|
updated_state["offer"] = offer
|
|
# 回答生成完成；失败收尾后的说明保持失败状态，不交付动作
|
|
if state.get("status") != "failed":
|
|
updated_state["status"] = "completed"
|
|
|
|
# 返回包含回答和完成状态的更新结果
|
|
return updated_state
|
|
|
|
async def _extract_offer(self, prompt: str, answer: str) -> str | None:
|
|
try:
|
|
result = await get_llm(ExtractedOffer).ainvoke(
|
|
[
|
|
SystemMessage(content=OFFER_EXTRACT_PROMPT),
|
|
HumanMessage(
|
|
content=json.dumps(
|
|
{"user": prompt, "assistant": answer}, ensure_ascii=False
|
|
)
|
|
),
|
|
]
|
|
)
|
|
except Exception as error:
|
|
# 回答已交付，提取失败只影响下一轮能否直接执行，不让本轮失败
|
|
self.logger.warning(f"提取回答中的提议失败: {error!r}")
|
|
return None
|
|
return result.offer
|
|
|
|
@staticmethod
|
|
async def _stream(content: str, emit) -> str:
|
|
# 获取普通对话模型，问答只需文本回答，不绑定结构化输出协议
|
|
llm_client = get_llm()
|
|
parser = StrOutputParser()
|
|
parts: list[str] = []
|
|
# 异步流式调用模型，结合用户问题、当前大屏和澄清记录生成回答
|
|
async for chunk in llm_client.astream(
|
|
[
|
|
# 系统消息提供稳定的问答规则和回答边界
|
|
SystemMessage(content=QA_SYSTEM_PROMPT),
|
|
# 输入上下文序列化为 JSON，只保留显式设置的字段
|
|
HumanMessage(content=content),
|
|
]
|
|
):
|
|
# 从每个片段提取文本，兼容文本内容块，工具调用等无文本片段会得到空串
|
|
delta = parser.invoke(chunk)
|
|
if delta:
|
|
parts.append(delta)
|
|
emit(delta)
|
|
# 拼接完整回答，并去除首尾的空白
|
|
return "".join(parts).strip()
|