group_xinghuo_jinrong/app/service/metric_ambiguity.py

"""口径字典歧义检测（D-11 · 沉淀前 / 问数 clarify 共用结构）。"""
from __future__ import annotations

import json
from typing import Any

from app.service.dict_service import Ambiguity, Metric, MetricRegistry, merge_registries, registry_with_published
from app.service.llm import DeepSeekLLM

DICT_CHECK_SYSTEM = (
    "你是金融数据口径审核助手。根据已有口径与新指标草稿，判断用户说法是否会混淆。"
    "只输出 JSON：{\"summary\":\"一句话建议\",\"actions\":[\"建议1\",\"建议2\"]}"
    "不要 markdown。"
)


def _parse_aliases(raw: Any) -> list[str]:
    if raw is None:
        return []
    if isinstance(raw, list):
        return [str(x).strip() for x in raw if str(x).strip()]
    s = str(raw).strip()
    if not s:
        return []
    if s.startswith("["):
        try:
            arr = json.loads(s)
            if isinstance(arr, list):
                return [str(x).strip() for x in arr if str(x).strip()]
        except json.JSONDecodeError:
            pass
    return [p.strip() for p in s.replace("，", ",").split(",") if p.strip()]


def _issue_from_ambiguity(term: str, amb: Ambiguity) -> dict[str, Any]:
    return {
        "term": term,
        "options": [
            {
                "metric_key": c.key,
                "metric_name": c.name,
                "definition": c.definition,
            }
            for c in amb.candidates
        ],
    }


def detect_dict_ambiguity(
    *,
    metric_key: str,
    metric_name: str,
    definition: str,
    aliases: Any,
    repo: Any | None = None,
    base_registry: MetricRegistry | None = None,
    llm: DeepSeekLLM | None = None,
) -> dict[str, Any]:
    """合并草稿进 registry 后扫描别名/名称是否触发消歧。"""
    if base_registry is not None:
        reg = base_registry
    else:
        from app.service.analytics_repo import AnalyticsRepo

        reg = registry_with_published(repo or AnalyticsRepo())

    alias_list = _parse_aliases(aliases)
    draft = Metric(
        key=(metric_key or "").strip(),
        name=(metric_name or "").strip(),
        aliases=alias_list,
        definition=(definition or "").strip(),
    )
    merged = merge_registries(reg, [draft]) if draft.key and draft.name else reg

    terms: list[str] = []
    if draft.name:
        terms.append(draft.name)
    terms.extend(alias_list)

    issues: list[dict[str, Any]] = []
    seen: set[str] = set()
    for term in terms:
        if not term or term in seen:
            continue
        seen.add(term)
        resolved = merged.resolve(term)
        if isinstance(resolved, Ambiguity):
            issues.append(_issue_from_ambiguity(term, resolved))

    llm_summary = ""
    llm_actions: list[str] = []
    if issues and llm is not None:
        try:
            payload = {
                "existing": [m.to_dict() for m in reg.all()],
                "draft": draft.to_dict(),
                "conflicts": issues,
            }
            text, _ = llm.complete(
                [
                    {"role": "system", "content": DICT_CHECK_SYSTEM},
                    {
                        "role": "user",
                        "content": f"数据：{json.dumps(payload, ensure_ascii=False)[:6000]}\n请输出 JSON：",
                    },
                ],
                temperature=0.1,
                max_tokens=400,
            )
            data = json.loads(text.strip().strip("`").replace("json", "", 1).strip())
            llm_summary = str(data.get("summary") or "")
            raw_actions = data.get("actions") or []
            if isinstance(raw_actions, list):
                llm_actions = [str(a) for a in raw_actions[:5]]
        except Exception:
            llm_summary = "存在与现有口径重叠的说法，请收窄别名或改用更具体的指标名。"

    return {
        "needs_choice": bool(issues),
        "issues": issues,
        "llm_summary": llm_summary,
        "llm_actions": llm_actions,
    }


def clarify_payload_from_ambiguity(amb: Ambiguity, *, llm_hint: str | None = None) -> dict[str, Any]:
    return {
        "term": amb.term,
        "options": [
            {"metric_key": c.key, "metric_name": c.name, "definition": c.definition}
            for c in amb.candidates
        ],
        "llm_hint": llm_hint,
    }