"""自然语言查询服务。

采用"结构化意图优先 + LLM 兜底"策略：
- 若问题命中线索检索意图（置信度/场景/状态/列出线索等），直接查审计数据库返回真实结果，
  实现"数据找人"，不依赖外部模型，数据不出域。
- 其余开放性问题再交给 LLMProvider（本地优先）。
对应 R4 / R20 / R7。
"""

from __future__ import annotations

from dataclasses import dataclass

from sqlalchemy.orm import Session

from app.clues import service as clue_svc
from app.clues.models import ClueStatus, ConfidenceTier
from app.llm import ChatMessage, get_llm_provider

SYSTEM_PROMPT = (
    "你是电信运营商内部审计助手。基于审计数据中台的数据回答问题，"
    "给出可解释的依据；无证据支撑时明确说明，不臆造数据。"
)

# 关键词 → 过滤条件映射
_CONFIDENCE_KW = {"高置信": ConfidenceTier.HIGH, "高风险": ConfidenceTier.HIGH,
                  "中置信": ConfidenceTier.MEDIUM, "低置信": ConfidenceTier.LOW}
_SCENARIO_KW = {"拆单": "R8", "政企": "R8", "养卡": "R9", "骗补": "R9", "彩铃": "R9"}
_STATUS_KW = {"待处理": ClueStatus.NEW, "已分派": ClueStatus.ASSIGNED,
              "研判": ClueStatus.REVIEWING, "属实": ClueStatus.CONFIRMED,
              "误报": ClueStatus.DISMISSED, "已销项": ClueStatus.CLOSED}
_LIST_KW = ("线索", "列出", "查", "有哪些", "多少", "列表", "看看", "显示")

_SCENARIO_NAME = {"R8": "政企拆单", "R9": "养卡骗补"}
_CONF_NAME = {ConfidenceTier.HIGH: "高置信", ConfidenceTier.MEDIUM: "中置信",
              ConfidenceTier.LOW: "低置信"}


@dataclass
class NLQAnswer:
    question: str
    answer: str
    provider: str
    model: str
    egress: bool


def _match_first(question: str, mapping: dict):
    for kw, val in mapping.items():
        if kw in question:
            return val
    return None


def _is_clue_query(question: str) -> bool:
    return any(kw in question for kw in _LIST_KW) or any(
        kw in question for kw in {**_CONFIDENCE_KW, **_SCENARIO_KW, **_STATUS_KW}
    )


def _format_clue_answer(question: str, clues: list) -> str:
    if not clues:
        return "未检索到符合条件的线索。可调整筛选条件，或先运行扫描生成线索。"
    lines = [f"共检索到 {len(clues)} 条线索："]
    for i, c in enumerate(clues, 1):
        amount = f"，涉及金额约 {c.amount_involved/10000:.1f} 万元" if c.amount_involved else ""
        lines.append(
            f"{i}. [{_SCENARIO_NAME.get(c.scenario_code, c.scenario_code)}] {c.title}"
            f"（{_CONF_NAME.get(c.confidence, c.confidence.value)}，评分 {c.score:.2f}{amount}）"
            f"——{c.rationale}"
        )
    return "\n".join(lines)


def ask(question: str, session: Session | None = None) -> NLQAnswer:
    """处理一次自然语言查询：优先结构化检索，其余交给 LLM。"""
    # 结构化意图：检索线索（数据找人，不出域）
    if session is not None and _is_clue_query(question):
        confidence = _match_first(question, _CONFIDENCE_KW)
        scenario = _match_first(question, _SCENARIO_KW)
        status = _match_first(question, _STATUS_KW)
        clues = clue_svc.list_clues(
            session, status=status, scenario_code=scenario, confidence=confidence
        )
        return NLQAnswer(
            question=question,
            answer=_format_clue_answer(question, clues),
            provider="datahub",
            model="结构化检索",
            egress=False,
        )

    # 开放性问题：交给 LLM（本地优先）
    provider = get_llm_provider()
    messages = [
        ChatMessage(role="system", content=SYSTEM_PROMPT),
        ChatMessage(role="user", content=question),
    ]
    resp = provider.chat(messages)
    return NLQAnswer(
        question=question,
        answer=resp.content,
        provider=resp.provider,
        model=resp.model,
        egress=resp.egress,
    )