Create-knowledge-find-decode/scripts/build_creation_demo.py

"""创作知识 query 正交 demo：5 套家族机械正交 → LLM 只做排除(query_filter.txt) → 存 JSON。

不真实搜，只产 query 供前端看。轴严格取分类树的"创作支"：
  实质 = 实质树·理念支(排除表象)   形式 = 形式树·架构支(排除呈现)
  目的池 = 作用树 + 感受树 + 意图树 全部合一(随机取)
  阶段意图 = 灵感/选题/脚本 展开成创作者真会搜的词(选题/开头/钩子/标题/封面/文案…)——脊柱，每族必带
  模态 = 视频/图片   知识类型 = 怎么做/有哪些/为什么
脊柱(每条都带)：… 阶段意图 + 模态 + 知识类型；前面配 实质/形式/目的 之一或组合。
5 家族：① 实质×阶段 ② 形式×阶段 ③ 实质×形式×阶段 ④ 目的池×阶段 ⑤ 纯阶段，各 20 条。
每条原串过 query_filter.txt(keep/排除)，存 keep+reason 供前端展示。
用法：PYTHONPATH=. CK_ENV_FILE=.env python scripts/build_creation_demo.py
"""
from __future__ import annotations

import json
import random
from pathlib import Path

import httpx

from core.config import Settings

ROOT = Path(__file__).resolve().parent.parent
TREES = ROOT / "scope_trees" / "trees_index.json"
FILTER_PROMPT = ROOT / "acquisition" / "query_filter.txt"   # 筛选词在 acquisition/
OUT = ROOT / "data" / "queries" / "creation_demo.json"
PER = 20
KTYPE = ["怎么做", "有哪些", "为什么"]
MODALITY = ["视频", "图片"]   # 被创作内容的形态(与教学帖本身格式无关),正交进所有家族
# 创作阶段意图轴(脊柱):灵感/选题/脚本 展开成创作者真会搜的词(query构造.md)。真实数据里
# "脚本"0次、但 开头/钩子/标题/封面/选题 各几十次——故用展开词,不用三个干阶段词。
STAGE_INTENT = {
    "灵感": ["找素材", "内容方向", "案例拆解", "灵感"],
    "选题": ["选题", "爆款选题", "选题方向"],
    "脚本": ["脚本", "文案", "开头", "钩子", "结构", "标题", "封面", "结尾"],
}
INTENT = [w for ws in STAGE_INTENT.values() for w in ws]   # 扁平成一个池,随机取


def _segs(p):
    return [x for x in (p or "").split("/") if x]


def _leaves(idx, source_type, under=None):
    """某树某支下的叶子节点名(没有更深子节点的=元素层)。under 限定分支。"""
    paths = [(_segs(n["path"]), n.get("name")) for n in idx if n.get("source_type") == source_type]
    if under:
        paths = [(s, nm) for s, nm in paths if under in s]
    allp = {"/".join(s) for s, _ in paths}
    out, seen = [], set()
    for s, nm in paths:
        if len(s) < 2:
            continue
        full = "/".join(s)
        is_leaf = not any(o != full and o.startswith(full + "/") for o in allp)
        name = nm or s[-1]
        if is_leaf and name and name not in seen:
            seen.add(name)
            out.append(name)
    return out


def _nonleaf(idx, source_type, depths=(3, 4), under=None):
    """某树某支下、指定层级的【非叶子"类目"节点】(底下还有元素,不取元素本身)。
    对齐制作侧取法:实质/形式 取 depth 3-4 的类目层,而非最深的元素层。"""
    paths = [(_segs(n["path"]), n.get("name")) for n in idx if n.get("source_type") == source_type]
    if under:
        paths = [(s, nm) for s, nm in paths if under in s]
    allp = {"/".join(s) for s, _ in paths}
    out, seen = [], set()
    for s, nm in paths:
        if len(s) not in depths:
            continue
        full = "/".join(s)
        is_nonleaf = any(o != full and o.startswith(full + "/") for o in allp)
        name = nm or (s[-1] if s else "")
        if is_nonleaf and name and name not in seen:
            seen.add(name)
            out.append(name)
    return out


def _filter(queries, settings):
    """把一批原串喂 query_filter.txt(LLM 只做 keep/排除)。返回 [{keep,reason}] 对齐顺序。"""
    user = json.dumps([{"idx": i, "query": q} for i, q in enumerate(queries)], ensure_ascii=False)
    api = settings.openrouter_base_url.rstrip("/") + "/chat/completions"
    headers = {"Authorization": f"Bearer {settings.openrouter_api_key}", "Content-Type": "application/json"}
    body = {"model": settings.llm_model, "messages": [
        {"role": "system", "content": FILTER_PROMPT.read_text("utf-8")},
        {"role": "user", "content": user}], "response_format": {"type": "json_object"}}
    try:
        resp = httpx.post(api, headers=headers, json=body, timeout=120)
        resp.raise_for_status()
        txt = resp.json()["choices"][0]["message"]["content"]
        # query_filter 要求输出数组；有的模型会包一层 {"result":[...]}，都兜住
        data = json.loads(txt)
        arr = data if isinstance(data, list) else next((v for v in data.values() if isinstance(v, list)), [])
        by = {d.get("idx"): d for d in arr if isinstance(d, dict)}
        return [{"keep": bool(by.get(i, {}).get("keep", True)),
                 "reason": str(by.get(i, {}).get("reason", ""))[:50]} for i in range(len(queries))]
    except Exception as exc:
        return [{"keep": True, "reason": f"筛选失败:{str(exc)[:30]}"} for _ in queries]


def main():
    settings = Settings.from_env()
    rng = random.Random(7)
    idx = json.loads(TREES.read_text("utf-8"))
    SHI = _nonleaf(idx, "实质", depths=(3, 4), under="理念")   # 类目层,非元素
    XING = _nonleaf(idx, "形式", depths=(3, 4), under="架构")  # 类目层,非元素
    POOL = _leaves(idx, "作用") + _leaves(idx, "感受") + _leaves(idx, "意图")
    print(f"实质 {len(SHI)} / 形式 {len(XING)} / 目的池 {len(POOL)} / 业务阶段 {len(INTENT)}")

    def pick(seq):
        return rng.choice(seq)

    # 每家族：生成器 + 用到的轴(给前端标列)
    # 业务阶段=脊柱,每族必带;模态+知识类型固定收尾;前面配 实质/形式/目的 之一或组合
    families = [
        {"key": "f1", "name": "实质 × 业务阶段", "axes": ["实质", "模态", "业务阶段", "知识类型"],
         "gen": lambda: {"parts": {"实质": pick(SHI), "业务阶段": pick(INTENT), "模态": pick(MODALITY), "知识类型": pick(KTYPE)}}},
        {"key": "f2", "name": "形式 × 业务阶段", "axes": ["形式", "模态", "业务阶段", "知识类型"],
         "gen": lambda: {"parts": {"形式": pick(XING), "业务阶段": pick(INTENT), "模态": pick(MODALITY), "知识类型": pick(KTYPE)}}},
        {"key": "f3", "name": "实质 × 形式 × 业务阶段", "axes": ["实质", "形式", "模态", "业务阶段", "知识类型"],
         "gen": lambda: {"parts": {"实质": pick(SHI), "形式": pick(XING), "业务阶段": pick(INTENT), "模态": pick(MODALITY), "知识类型": pick(KTYPE)}}},
        {"key": "f4", "name": "(作用/感受/意图) × 业务阶段", "axes": ["作用/感受/意图", "模态", "业务阶段", "知识类型"],
         "gen": lambda: {"parts": {"目的": pick(POOL), "业务阶段": pick(INTENT), "模态": pick(MODALITY), "知识类型": pick(KTYPE)}}},
        {"key": "f5", "name": "纯业务阶段", "axes": ["模态", "业务阶段", "知识类型"],
         "gen": lambda: {"parts": {"业务阶段": pick(INTENT), "模态": pick(MODALITY), "知识类型": pick(KTYPE)}}},
    ]
    # 各部件按固定顺序拼成原串（内容维度在前，模态贴题材后，业务阶段+知识类型收尾）
    order = ["实质", "形式", "目的", "模态", "业务阶段", "知识类型"]

    # 业务阶段值存「分组结构」(STAGE_INTENT)，前端按 灵感/选题/脚本 分组+缩进展示；其余轴是扁平数组
    out = {"axis_values": {"实质": SHI, "形式": XING, "目的池": POOL, "业务阶段": STAGE_INTENT, "模态": MODALITY, "知识类型": KTYPE},
           "families": []}
    for fam in families:
        seen, items = set(), []
        while len(items) < PER and len(seen) < PER * 40:
            parts = fam["gen"]()["parts"]
            q = " ".join(parts[k] for k in order if k in parts)
            if q in seen:
                continue
            seen.add(q)
            items.append({"query": q, "parts": parts})
        verdicts = _filter([it["query"] for it in items], settings)
        for it, v in zip(items, verdicts):
            it.update(v)
        kept = sum(1 for it in items if it["keep"])
        print(f"[{fam['name']}] 生成 {len(items)} 条, 筛后保留 {kept}")
        out["families"].append({"key": fam["key"], "name": fam["name"], "axes": fam["axes"], "items": items})

    OUT.parent.mkdir(parents=True, exist_ok=True)
    OUT.write_text(json.dumps(out, ensure_ascii=False, indent=1), encoding="utf-8")
    print(f"→ {OUT}")


if __name__ == "__main__":
    main()