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 = """# 任务：投顾助手意图识别
你是投顾系统意图分类器，对客户输入文本做意图识别，**只能输出一个类别名称，禁止额外解释**。

## 类别说明
1. 基金推荐：客户希望推荐、筛选基金产品
2. 调仓建议：客户询问是否买卖、加减仓、更换基金，寻求调仓交易建议
3. 客户持仓基金分析：分析客户现有持仓组合、风险、收益情况，不涉及买卖操作建议
4. 跟客户的沟通话术：投顾需要生成一段发给客户的话术文案。【注意：客户发起提问，不会是该类别】
5. 数据查询：查询基金客观数据，如净值、基金经理、规模、持仓、费率等事实信息，不做推荐、诊断
6. 普通聊天：日常问候、单纯情绪吐槽，无明确业务诉求

## 判定优先级（同时存在多个诉求时，取优先级最高）
调仓建议 > 基金推荐 > 客户持仓基金分析 > 数据查询 > 跟客户的沟通话术 > 普通聊天

## 示例(xx表示我名下的客户名字)
输入：帮我名下XX推荐几只适合养老的基金
输出：基金推荐

输入：XX手上的某某基金现在要不要卖出？
输出：调仓建议

输入：帮我看看xx的基金组合风险高不高
输出：客户持仓基金分析

输入：帮我查一下XX基金最新规模
输出：数据查询

输入：帮我写一段话安抚客户，解释近期回撤
输出：跟客户的沟通话术

输入：今天天气不错
输出：普通聊天

输入：xx持有的这几只基金波动很大，要不要减仓？
输出：调仓建议

输入：帮我看下xx持仓的基金，它们的基金经理是谁
输出：数据查询

现在开始分类
输入：{{user_query}}
输出：

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")
_REBALANCE_TERMS = ("调仓", "再平衡", "组合调整", "配置偏离", "偏离目标")
_DIALOGUE_TERMS = ("话术", "沟通", "怎么跟客户说", "如何向客户解释", "安抚客户", "投诉处理")
_CUSTOMER_DATA_TERMS = ("持仓", "资产", "余额", "交易", "账户", "份额", "市值", "盈亏", "资金")
_REFERENCE_TERMS = ("他", "她", "它", "这个客户", "该客户", "那个客户", "刚才", "上一轮")
_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"),
    )


def resolve_advisor_intent(
    query: str | None,
    classified_intent: str,
    *,
    customer_resolved: bool,
) -> str:
    """结合实体解析结果做最终业务路由，处理一句话中的明确优先意图。"""
    text = re.sub(r"\s+", "", query or "")
    if any(term in text for term in _REBALANCE_TERMS):
        return AGENT_INTENT_REBALANCE
    if any(term in text for term in _RECOMMENDATION_TERMS):
        return AGENT_INTENT_RECOMMEND
    if any(term in text for term in _DIALOGUE_TERMS):
        return AGENT_INTENT_DIALOGUE_SCRIPT
    if customer_resolved and any(term in text for term in _CUSTOMER_DATA_TERMS):
        return AGENT_INTENT_DATA_QUERY
    return classified_intent


async def classify_advisor_intent(
    query: str | None,
    llm_client=None,
    *,
    explicit_intent: str | None = None,
    conversation_context: str = "",
    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",
        )
    normalized_query = re.sub(r"\s+", "", query)
    if (
        any(term in normalized_query for term in _CUSTOMER_DATA_TERMS)
        and conversation_context
        and any(term in normalized_query for term in _REFERENCE_TERMS)
    ):
        return IntentClassification(
            AGENT_INTENT_DATA_QUERY,
            0.9,
            "rule_context",
            "上下文中的客户数据追问",
        )
    if llm_client is not None:
        try:
            prompt = query.strip()
            if conversation_context:
                prompt = f"会话上下文：\n{conversation_context[:2000]}\n\n当前问题：\n{prompt}"
            raw = await asyncio.wait_for(
                llm_client.chat(
                    [
                        {"role": "system", "content": _CLASSIFIER_PROMPT},
                        {"role": "user", "content": prompt},
                    ],
                    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", "resolve_advisor_intent"]