"""创作知识 query 正交 demo：多种正交组合 → LLM 评 valid/relevant(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) 业务阶段 = 灵感/选题/脚本(只这三个裸阶段，不展开)——脊柱，每族必带 模态 = 视频/图文 知识类型 = 怎么做/有哪些/为什么 新尾缀：模态 × 业务阶段 × 知识类型；老尾缀对照：阶段 × 动作 × [作用] × 知识类型。 实质 / 形式 / 目的等轴从同一张 MASTER 主表投影，方便控制变量横向对比。 每条原串过 query_filter.txt(valid + relevant)，存 keep+valid+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 import sys from pathlib import Path from acquisition.query import ACTIONS, _nonleaf_d4 # 老正交动作轴 + 4级非叶子取作用节点 from acquisition.query import STAGES as OLD_STAGES # 老的裸阶段 灵感/选题/脚本 from acquisition.query_filter import filter_queries # 共享筛选器(query_filter.txt) from core.config import Settings ROOT = Path(__file__).resolve().parent.parent TREES = ROOT / "scope_trees" / "trees_index.json" OUT = ROOT / "data" / "queries" / "creation_demo.json" PER = 30 BATCH_N = 30 # 全 demo 统一抽这么多个「实质 / 形式」，各族共用同一批；PER=BATCH_N 保证整批都用上、左右一一对应 KTYPE = ["怎么做", "有哪些", "为什么"] MODALITY = ["视频", "图文"] # 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 # 业务阶段(脊柱)：只留 3 个裸阶段，不再展开成 query构造.md 的衍生词(找素材/开头/钩子/标题…全去掉) INTENT = ["灵感", "选题", "脚本"] 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 main(): settings = Settings.from_env() rng = random.Random(7) DRY = "--dry" in sys.argv # 干跑：跳过 LLM 筛选(全 keep)，只验证生成结构/控制变量 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)}") # 统一批：30 实质 / 30 形式 / 30 目的(作用感受意图)，全 demo 共用 SHI_BATCH = rng.sample(SHI, min(BATCH_N, len(SHI))) XING_BATCH = rng.sample(XING, min(BATCH_N, len(XING))) POOL_BATCH = rng.sample(POOL, min(BATCH_N, len(POOL))) F6_ZY = _nonleaf_d4("作用", 10) # 老正交的"作用"(4级非叶子) F6_STAGE_ACT = [(s, a) for s in OLD_STAGES for a in ACTIONS] + [("", "")] # 阶段×动作 + 无动作变体 print(f"统一批: 实质×{len(SHI_BATCH)} 形式×{len(XING_BATCH)} 目的×{len(POOL_BATCH)}") # 主表：第 i 行把【所有轴】取值一次定死；各组合方式只是从这行挑自己用到的轴(投影)，严格控制变量 # ——同一个实质(在第 i 行)无论出现在哪种组合里，配的模态/业务阶段/知识类型都相同。 MASTER = [] for i in range(PER): st, ac = F6_STAGE_ACT[i % len(F6_STAGE_ACT)] MASTER.append({ "实质": SHI_BATCH[i % len(SHI_BATCH)], "形式": XING_BATCH[i % len(XING_BATCH)], "目的": POOL_BATCH[i % len(POOL_BATCH)], "模态": MODALITY[i % len(MODALITY)], "业务阶段": INTENT[i % len(INTENT)], "知识类型": KTYPE[(i // 3) % len(KTYPE)], # 与业务阶段解耦，纯尾缀族也能多出几种 "阶段": st or "/", "动作": ac or "/", "_st": st, "_ac": ac, "作用": F6_ZY[i % len(F6_ZY)], }) def proj(i, keys): # 新脊柱族：从主表第 i 行取这些 parts 键(query 由 order 拼) row = MASTER[i] return {"parts": {k: row[k] for k in keys}} def gen_old(i, *, shi=False, xing=False, purpose=False, zy=True): # 老正交族：[实质] [形式] [目的] (阶段+动作) [作用] 知识类型 r = MASTER[i] seg = (r["_st"] + r["_ac"]) if r["_ac"] else "" # 脚本撰写 / 空 head = ([r["实质"]] if shi else []) + ([r["形式"]] if xing else []) + ([r["目的"]] if purpose else []) tail = ([r["作用"]] if zy else []) + [r["知识类型"]] q = " ".join(head + ([seg] if seg else []) + tail) parts = {} if shi: parts["实质"] = r["实质"] if xing: parts["形式"] = r["形式"] if purpose: parts["目的"] = r["目的"] parts["阶段"], parts["动作"] = r["阶段"], r["动作"] if zy: parts["作用"] = r["作用"] parts["知识类型"] = r["知识类型"] return {"parts": parts, "query": q} # 每种组合方式：生成器 + 用到的轴。name = axes 用「×」连接，直接反映正交结构。 # axes 顺序即前端列顺序；前 5 种使用新尾缀，后 6 种用于对照老尾缀。 families = [ {"key": "f1", "axes": ["实质", "模态", "业务阶段", "知识类型"], "gen": lambda i: proj(i, ["实质", "模态", "业务阶段", "知识类型"])}, {"key": "f2", "axes": ["形式", "模态", "业务阶段", "知识类型"], "gen": lambda i: proj(i, ["形式", "模态", "业务阶段", "知识类型"])}, {"key": "f4", "axes": ["作用/感受/意图", "模态", "业务阶段", "知识类型"], "gen": lambda i: proj(i, ["目的", "模态", "业务阶段", "知识类型"])}, {"key": "f3", "axes": ["实质", "形式", "模态", "业务阶段", "知识类型"], "gen": lambda i: proj(i, ["实质", "形式", "模态", "业务阶段", "知识类型"])}, {"key": "f5", "axes": ["模态", "业务阶段", "知识类型"], "gen": lambda i: proj(i, ["模态", "业务阶段", "知识类型"])}, # 老正交·尾缀A：阶段 × 动作 × 作用 × 知识类型(五种正交) {"key": "a_shi", "axes": ["实质", "阶段", "动作", "作用", "知识类型"], "gen": lambda i: gen_old(i, shi=True, zy=True)}, {"key": "a_xing", "axes": ["形式", "阶段", "动作", "作用", "知识类型"], "gen": lambda i: gen_old(i, xing=True, zy=True)}, {"key": "a_both", "axes": ["实质", "形式", "阶段", "动作", "作用", "知识类型"], "gen": lambda i: gen_old(i, shi=True, xing=True, zy=True)}, {"key": "a_purpose", "axes": ["作用/感受/意图", "阶段", "动作", "作用", "知识类型"], "gen": lambda i: gen_old(i, purpose=True, zy=True)}, {"key": "a_tail", "axes": ["阶段", "动作", "作用", "知识类型"], "gen": lambda i: gen_old(i, zy=True)}, # 老正交·尾缀B：阶段 × 动作 × 知识类型(抽掉作用，五种正交) {"key": "b_shi", "axes": ["实质", "阶段", "动作", "知识类型"], "gen": lambda i: gen_old(i, shi=True, zy=False)}, {"key": "b_xing", "axes": ["形式", "阶段", "动作", "知识类型"], "gen": lambda i: gen_old(i, xing=True, zy=False)}, {"key": "b_both", "axes": ["实质", "形式", "阶段", "动作", "知识类型"], "gen": lambda i: gen_old(i, shi=True, xing=True, zy=False)}, {"key": "b_purpose", "axes": ["作用/感受/意图", "阶段", "动作", "知识类型"], "gen": lambda i: gen_old(i, purpose=True, zy=False)}, {"key": "b_tail", "axes": ["阶段", "动作", "知识类型"], "gen": lambda i: gen_old(i, zy=False)}, ] # 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"] # 业务阶段只剩 灵感/选题/脚本 三个，扁平数组(前端按池子平铺，不再分组缩进) out = {"axis_values": {"实质": SHI, "形式": XING, "目的池": POOL, "业务阶段": INTENT, "模态": MODALITY, "知识类型": KTYPE, "阶段": OLD_STAGES, "动作": ACTIONS, "作用": F6_ZY}, # 老尾缀对照轴 "families": []} for fam in families: name = " × ".join(fam["axes"]) # 家族名直接反映正交结构 seen, items = set(), [] for i in range(PER): # 单趟过主表、去重(纯尾缀族会自然少于 PER) g = fam["gen"](i) parts = g["parts"] q = g.get("query") or " ".join(parts[k] for k in order if k in parts) # f6/f7 自带 query 串 if q in seen: continue seen.add(q) items.append({"query": q, "parts": parts}) verdicts = ([{"keep": True, "valid": None, "relevant": True, "reason": "(dry:未过筛)"} for _ in items] if DRY else filter_queries([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"[{name}] 生成 {len(items)} 条, 筛后保留 {kept}") out["families"].append({"key": fam["key"], "name": 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()