Mutual_Fund/agent/advisor_agent/intent/classifier.py

"""投顾聊天意图分类：LLM 优先，规则识别兜底。"""
from __future__ import annotations

import asyncio
import json
import re
from dataclasses import dataclass

from common.common_const import (
    AGENT_INTENT_CASUAL_CHAT,
    AGENT_INTENT_DATA_QUERY,
    AGENT_INTENT_DIALOGUE_SCRIPT,
    AGENT_INTENT_FUND_ANALYSIS,
    AGENT_INTENT_REBALANCE,
    AGENT_INTENT_RECOMMEND,
)
from agent.advisor_agent.intent.recognizer import recognize_advisor_intent


VALID_INTENTS = frozenset({
    AGENT_INTENT_RECOMMEND,
    AGENT_INTENT_REBALANCE,
    AGENT_INTENT_FUND_ANALYSIS,
    AGENT_INTENT_DIALOGUE_SCRIPT,
    AGENT_INTENT_DATA_QUERY,
    AGENT_INTENT_CASUAL_CHAT,
})


@dataclass(frozen=True)
class IntentClassification:
    intent: str
    confidence: float
    source: str
    reason: str = ""


_CLASSIFIER_PROMPT = """你是基金投顾工作台的意图分类器，只负责分类，不回答用户问题。
只能从以下分类中选择一个：
- recommend：基金推荐、组合配置或投资方案
  例：“帮我推荐两只适合稳健型客户的基金”“给客户生成一份组合配置建议”“该买什么基金”
- rebalance：组合偏离、调仓、再平衡、仓位调整
  例：“该客户组合偏离目标配置，请给出调仓方案”“组合需要再平衡，降低股票类仓位”
- fund_analysis：单只基金分析、净值、收益、回撤、波动率、夏普比率
  例：“分析华夏回报近一年的净值走势和最大回撤”“这只基金的夏普比率和波动率如何”
- dialogue-script：给客户准备沟通话术、解释、安抚、投诉或风险提醒
  例：“市场波动时怎么和客户解释”“帮我准备安抚客户的沟通话术”“客户投诉了，话术怎么准备”
- data_query：查询客户持仓、资产、余额、收益、交易、账户明细、客户名册
  例：“查询客户48当前持仓和账户余额”“我名下有哪些客户”“统计名下客户数量”
- casual_chat：问候、闲聊、感谢、身份询问或无法归入业务分类的内容
  例：“你好”“谢谢”“你是谁”

只输出 JSON，不要 Markdown，不要额外文字：
{"intent":"分类值","confidence":0到1之间的数字,"reason":"不超过30字的原因"}
"""

_RECOMMENDATION_TERMS = ("推荐", "组合建议", "配置建议", "买什么", "适合配置", "筛选基金", "投资方案")
_SCOPE_ERROR_MESSAGE = "投顾范围查询仅支持客户数据查询"


def _rule_fallback(query: str | None) -> IntentClassification:
    intent = recognize_advisor_intent(query)
    if intent:
        return IntentClassification(intent=intent, confidence=0.72, source="rule", reason="关键词规则匹配")
    return IntentClassification(intent=AGENT_INTENT_CASUAL_CHAT, confidence=0.0, source="fallback", reason="无法匹配业务意图")


def _parse_model_result(raw: str) -> IntentClassification | None:
    text = raw.strip()
    fenced = re.search(r"\{.*\}", text, re.DOTALL)
    if fenced:
        text = fenced.group(0)
    try:
        payload = json.loads(text)
    except (TypeError, json.JSONDecodeError):
        return None
    intent = payload.get("intent")
    if intent not in VALID_INTENTS:
        return None
    try:
        confidence = max(0.0, min(1.0, float(payload.get("confidence", 0.0))))
    except (TypeError, ValueError):
        confidence = 0.0
    return IntentClassification(
        intent=intent,
        confidence=confidence,
        source="llm",
        reason=str(payload.get("reason") or "模型分类"),
    )


async def classify_advisor_intent(
    query: str | None,
    llm_client=None,
    *,
    explicit_intent: str | None = None,
    timeout: float = 2.0,
) -> IntentClassification:
    """分类用户意图；显式意图兼容旧客户端，模型失败时安全回退规则。"""
    if explicit_intent in VALID_INTENTS:
        return IntentClassification(explicit_intent, 1.0, "explicit", "客户端显式指定")
    if not query or not query.strip():
        return IntentClassification("", 0.0, "fallback", "空输入")
    if re.sub(r"\s+", "", query) == _SCOPE_ERROR_MESSAGE:
        return IntentClassification(
            AGENT_INTENT_CASUAL_CHAT,
            1.0,
            "rule",
            "识别为系统提示文本而非业务查询",
        )
    fast_intent = recognize_advisor_intent(query)
    if (
        fast_intent == AGENT_INTENT_DATA_QUERY
        and not any(term in query for term in _RECOMMENDATION_TERMS)
    ):
        return IntentClassification(
            fast_intent,
            0.95,
            "rule_fast",
            "明显查询类问题",
        )
    if llm_client is not None:
        try:
            raw = await asyncio.wait_for(
                llm_client.chat(
                    [
                        {"role": "system", "content": _CLASSIFIER_PROMPT},
                        {"role": "user", "content": query.strip()},
                    ],
                    temperature=0,
                    max_tokens=120,
                ),
                timeout=timeout,
            )
            parsed = _parse_model_result(raw)
            if parsed is not None:
                # LLM 偶尔会把“查询持仓/资产”等只读请求误判为推荐；
                # 对明确的查询动作以规则结果为准，避免误进入草稿生成分支。
                rule_intent = recognize_advisor_intent(query)
                if (
                    rule_intent == AGENT_INTENT_DATA_QUERY
                    and parsed.intent != AGENT_INTENT_DATA_QUERY
                    and not any(term in (query or "") for term in _RECOMMENDATION_TERMS)
                ):
                    return IntentClassification(
                        intent=rule_intent,
                        confidence=max(parsed.confidence, 0.9),
                        source="rule_override",
                        reason="明确查询类关键词覆盖模型误判",
                    )
                return parsed
        except Exception:
            pass
    return _rule_fallback(query)


__all__ = ["IntentClassification", "classify_advisor_intent"]