beige/backend/app/services/ai_engine/fission_fallback.py

"""
app/services/ai_engine/fission_fallback.py — 裂变品类兜底（从 fission_prompt 拆出）

LLM 挂/返回不可解析时，按品类生成完整草稿，保证不卡用户（对齐上线版 split.js
的 inferCategory / fallbackAnglesByCategory / buildFallbackNotes）。
拆分原因：fission_prompt.py 超 200 行红线上限，品类兜底是内聚可独立的一块。
"""
from __future__ import annotations

import re

from app.services.ai_engine.fission_prompt import (
    normalize_tags,
    sanitize_image_plan_text,
)

_CATEGORY_PATTERNS = [
    ("个护护理", re.compile(r"护手|手霜|身体乳|润肤|唇膏|洗护|护理")),
    ("美妆护肤", re.compile(r"霜|乳|精华|面膜|粉底|彩妆|口红|护肤|美妆")),
    ("食品饮品", re.compile(r"饮|茶|咖啡|果汁|奶|冲泡|零食|食品|吃")),
    ("营养健康", re.compile(r"维生素|益生菌|蛋白|营养|保健|膳食")),
    ("家居生活", re.compile(r"收纳|清洁|家居|厨房|小物|工具|电器")),
    ("服饰穿搭", re.compile(r"衣|裤|鞋|包|穿搭|面料|服饰")),
]


def infer_category(product: dict) -> str:
    """按产品名/卖点/关键词推断品类（对齐上线版 inferCategory）。"""
    p = product or {}
    text = (
        str(p.get("name", "")) + "、".join(p.get("selling_points", []) or [])
        + "、".join(p.get("keywords", []) or [])
    )
    for cat, pattern in _CATEGORY_PATTERNS:
        if pattern.search(text):
            return cat
    return "通用好物"


def _fallback_angles(category: str, product_name: str, points: list[str]) -> list[dict]:
    """按品类返回兜底角度（对齐上线版 fallbackAnglesByCategory，节选主品类+通用兜底）。"""
    name = product_name or "这个好物"
    p0 = points[0] if points else "到底好不好用"
    maps = {
        "个护护理": [
            {"dimension": "换人群", "title": f"{name}手干星人真的会回购！", "scene": "办公室/通勤", "painPoint": "手干、倒刺、涂完黏手"},
            {"dimension": "换场景", "title": "包里常备这支太省心了", "scene": "随身护理", "painPoint": "出门临时干到难受"},
            {"dimension": "换测评", "title": "不黏手这点太加分了！", "scene": "手部质地测评", "painPoint": "摸键盘手机都怕黏"},
            {"dimension": "换痛点", "title": "换季手粗糙别硬扛", "scene": "换季护理", "painPoint": "洗完手紧绷粗糙"},
            {"dimension": "换选择理由", "title": "这支属于会推荐给同事", "scene": "办公室分享", "painPoint": "想要清爽又好用"},
        ],
        "食品饮品": [
            {"dimension": "换场景", "title": "工位囤这个真的方便", "scene": "办公室饮用", "painPoint": "下午嘴馋又怕踩雷"},
            {"dimension": "换口感", "title": "第一口就知道没买错", "scene": "口感测评", "painPoint": "怕味道寡淡或太腻"},
            {"dimension": "换步骤", "title": "懒人冲泡也能很稳定", "scene": "快速准备", "painPoint": "想方便但不想牺牲口感"},
            {"dimension": "换囤货", "title": "这波囤在家里不心疼", "scene": "拆箱囤货", "painPoint": "高频喝/吃更看重性价比"},
            {"dimension": "换人群", "title": "打工人下午这口太需要了", "scene": "下午补给", "painPoint": "没精神但不想太复杂"},
        ],
    }
    return maps.get(category) or [
        {"dimension": "换人群", "title": f"{name}比想象中实用！", "scene": "真实使用", "painPoint": f"用户关心{p0}"},
        {"dimension": "换场景", "title": "这个场景下真的会用到", "scene": "日常场景", "painPoint": "买前不知道适不适合自己"},
        {"dimension": "换痛点", "title": "这个小问题终于被解决了", "scene": "痛点解决", "painPoint": "日常高频但容易被忽略的问题"},
        {"dimension": "换测评", "title": "细节近看才知道值不值", "scene": "细节测评", "painPoint": "怕宣传好看但实际一般"},
        {"dimension": "换转化", "title": "这波属于会推荐给朋友", "scene": "软性转化", "painPoint": "想要真实选择理由"},
    ]


_FALLBACK_OVERLAY = {
    "hook": "这也太自然了", "pain_scene": "这个痛点太真实",
    "applied_proof": "核心卖点看得见", "texture": "质地水润好推开",
    "social_proof": "身边人都在问", "scenario": "出门随手带",
    "tutorial": "三步快速出门", "closer": "这波真的会囤",
    "product_closeup": "单品细节近看", "ingredient": "成分看得见",
}
_FALLBACK_TEXT = {
    "applied_proof": "按当前品类生成核心证明页：用真实使用过程、细节近景、成分/口感/材质/质地等可感知证据证明卖点",
    "texture": "展示产品质地、材质、口感、成分或使用细节，让用户看到卖点证据",
    "closer": "拆箱、囤货角、通勤包或桌面场景，用省钱情报/暗号口吻做软性转化",
}


def build_fallback_image_plan(note: dict, image_count: int) -> list[dict]:
    """LLM挂时按叙事角色兜底 imagePlan（对齐上线版 buildFallbackImagePlan）。"""
    from app.services.ai_engine.storyboard import get_narrative_roles
    existing = note.get("imagePlan") if isinstance(note.get("imagePlan"), list) else []
    plan = []
    for i, role in enumerate(get_narrative_roles(image_count)):
        r = role.get("role", "")
        ex = existing[i] if i < len(existing) else {}
        plan.append({
            "role": r,
            "title": sanitize_image_plan_text(ex.get("title") or role.get("name", ""), 12),
            "overlayText": sanitize_image_plan_text(
                ex.get("overlayText") or note.get("coverTitle") or note.get("title")
                if r == "hook" else (ex.get("overlayText") or _FALLBACK_OVERLAY.get(r) or role.get("name", "")), 18),
            "text": sanitize_image_plan_text(
                ex.get("text") or _FALLBACK_TEXT.get(r) or role.get("focus", ""), 72),
        })
    return plan


def build_fallback_notes(
    source_note: dict, product: dict, note_count: int, image_count: int,
) -> list[dict]:
    """LLM返回不可解析时的品类兜底完整草稿（对齐上线版 buildFallbackNotes）。"""
    prod = product or {}
    src = source_note or {}
    name = prod.get("name", "") or "这款产品"
    points = prod.get("selling_points", []) or ["使用方便", "真实可感知", "适合日常场景", "性价比高"]
    audience = prod.get("target_audience", "") or "目标用户"
    keywords = prod.get("keywords", []) or []
    category = infer_category(prod)
    tags = normalize_tags(
        src.get("tags", []),
        keywords or [name, category, "真实测评", "好物分享"],
    )
    angles = _fallback_angles(category, name, points)
    kw = keywords or [x for x in [name, category, "真实测评", "好物分享"] if x]
    out = []
    for i in range(note_count):
        a = angles[i % len(angles)]
        main = points[i % len(points)]
        second = points[(i + 1) % len(points)]
        title = a["title"]
        note = {
            "title": title,
            "content": (
                f"姐妹们，这条先当真实测评草稿看。{name}主打{main}，对{audience}来说，"
                f"最有用的不是堆参数，而是解决「{a['painPoint']}」这个真实场景。✅\n\n"
                f"我会先看它在{a['scene']}里是不是真的方便，再看{second}有没有日常可感知的体验。✨ "
                f"如果不是那种一眼硬广的表达，反而更像朋友顺手分享。\n\n"
                f"如果你也在意{a['painPoint']}，这类选择理由会更容易判断适不适合自己。🌿"
            ),
            "tags": tags,
            "coverTitle": re.sub(r"[！!]", "", title),
            "dimension": a["dimension"],
            "audience": audience,
            "scene": a["scene"],
            "painPoint": a["painPoint"],
            "keywords": kw,
        }
        note["imagePlan"] = build_fallback_image_plan(note, image_count)
        out.append(note)
    return out