POD 趋势感知 Agent：缓存热点模式 + 三图合成 + 热点去重/风格去重 + review 兜底

- 缓存热点批量流程（有采集缓存不触发 Google） - 简报不足直接从采集缓存生成（轻量补齐） - 三图合成（模特/印花/底图）+ 底图压缩 <2MB - 热点去重→风格去重自动切换 + 不适合类目 review 兜底 - 透明背景（background=transparent）+ 提示词清洗（敏感词/背景描述） - 任务前 basemap 校验 + 模板国家校验 + 模特任务级分配
2026-08-22 14:14:01 +08:00
commit f493bde8a9
98 changed files with 10280 additions and 0 deletions
@@ -0,0 +1,127 @@
+"""节点 4/6：合规筛选（screen）。
+
+调用可插拔 LLM 后端做合规筛查 + 结构化四要素；后端调用失败时降级 MockBackend。
+按 topic 把筛查结果映射回 scored 候选（补 score/sources/country），再做最终风险过滤
+（blocked 丢弃；review 按 keep_review 决定）。
+"""
+from pathlib import Path
+from typing import Any, Dict, List
+
+from graph.loader import load_system_prompt
+from graph.llms import DEFAULT_SYSTEM_PROMPT, get_backend
+from graph.style_rules import COUNTRY_AESTHETICS
+from graph.validate import with_fallback
+
+
+@with_fallback("screen")
+def screen_node(state: Dict[str, Any]) -> Dict[str, Any]:
+    country = state["country"]
+    scored: List[Dict[str, Any]] = state.get("scored_rows") or []
+    config = state["config"]
+    cc = state["country_config"]
+    llm_cfg = config.get("llm_screen") or {}
+    provider = llm_cfg.get("provider", "mock")
+    keep_review = bool(llm_cfg.get("keep_review", False))
+    batch_size = int(llm_cfg.get("max_topics_per_call", 12))
+    blacklist = [str(b).lower() for b in (config.get("blacklist") or [])]
+
+    prompts_dir = Path(state["prompts_dir"])
+    system_prompt = load_system_prompt(prompts_dir, DEFAULT_SYSTEM_PROMPT)
+    aesthetic_hint = cc.get("style_hint") or COUNTRY_AESTHETICS.get(country, {}).get("style_hint", "")
+
+    # 排除已用热点（去重生效：已用 topic 不再进入本次简报，每次跑都用新热点）
+    try:
+        import json as _json
+        used_p = Path(state.get("cache_dir") or state.get("output_dir", "")) / "used_designs.json"
+        if used_p.exists():
+            ud = _json.loads(used_p.read_text(encoding="utf-8")).get("used", []) or []
+            used_topics = {str(u.get("topic", "")).strip().lower() for u in ud}
+            before = len(scored)
+            scored = [c for c in scored if str(c.get("topic", "")).strip().lower() not in used_topics]
+            if len(scored) < before:
+                print(f"[screen] 排除已用热点 {before - len(scored)} 条（去重），剩余 {len(scored)} 条可选")
+    except Exception:  # noqa: BLE001
+        pass
+
+    # 简报数量 = 用多少生成多少：llm_screen.max_briefs 配置优先，否则按扩展后的总任务数
+    #（每个产品一个热点；数量 N=每个款-颜色条目的设计数 → 总任务=条目数×N）；
+    # 前台显示不依赖简报（读 collected 完整池 + used_designs 剔除已用，用完即从前台消失）。
+    pcfg = config.get("product") or {}
+    ls_cfg = config.get("llm_screen") or {}
+    limit = int(ls_cfg.get("max_briefs") or 0)
+    if not limit and pcfg.get("spu_tasks"):
+        limit = len(pcfg.get("spu_tasks") or [])   # 扩展后总任务数（=产品数）
+    if not limit:
+        limit = int(pcfg.get("spu_count") or 0)
+    if limit > 0:
+        ordered = sorted(scored, key=lambda c: -(float(c.get("score") or 0)))
+        scored_limited = ordered[:limit]
+        print(f"[screen] 简报限量 {limit} → 筛前 {len(scored_limited)} 个高分热点（共 {len(scored)} 个）")
+    else:
+        scored_limited = scored
+        print(f"[screen] 简报全量 {len(scored_limited)} 条（未限量，全部生成简报）")
+
+    topics = [c["topic"] for c in scored_limited]
+
+    backend = get_backend(provider)
+    if provider != "mock":
+        backend.bind_config(llm_cfg)
+        if not backend.has_key:
+            print("[screen] 未检测到 LLM api_key（请配置 llm_screen.api_key 或环境变量 "
+                  "LLM_API_KEY/OPENAI_API_KEY），降级 Mock 兜底。")
+            backend = get_backend("mock")
+    try:
+        screened = backend.screen(topics, country, aesthetic_hint, system_prompt, blacklist, batch_size)
+    except Exception as e:  # noqa: BLE001
+        print(f"[screen] {provider} 调用失败，降级 Mock: {e}")
+        backend = get_backend("mock")
+        screened = backend.screen(topics, country, aesthetic_hint, system_prompt, blacklist, batch_size)
+
+    # 映射回 scored 候选
+    by_topic = {s.get("topic", "").lower(): s for s in screened}
+    out: List[Dict[str, Any]] = []
+    missing: List[Dict[str, Any]] = []
+    for it in scored_limited:
+        s = by_topic.get(it["topic"].lower())
+        if s is None:
+            missing.append(it)
+            continue
+        s = dict(s)
+        s["country"] = country
+        s["score"] = it.get("score", 0)
+        s["sources"] = it.get("sources", "")
+        out.append(s)
+
+    # LLM 漏判的主题用 Mock 单独补全，避免丢数据（而非整批降级）
+    if missing:
+        print(f"[screen] LLM 漏判 {len(missing)} 个主题，用 Mock 单独补全："
+              f"{[m['topic'] for m in missing]}")
+        mock = get_backend("mock")
+        m_res = mock.screen(
+            [m["topic"] for m in missing], country, aesthetic_hint,
+            system_prompt, blacklist, batch_size,
+        )
+        m_by = {r.get("topic", "").lower(): r for r in m_res}
+        for it in missing:
+            s = m_by.get(it["topic"].lower())
+            if s is None:
+                continue
+            s = dict(s)
+            s["country"] = country
+            s["score"] = it.get("score", 0)
+            s["sources"] = it.get("sources", "")
+            out.append(s)
+
+    # 最终风险过滤：只滤 blocked（硬拦截）；review（待复核）保留——
+    # 由 assign_hotspots 的 allow_review 决定是否参与分配（openai 模式 review+concept 可用），
+    # 避免 review 被静默丢弃导致"任务 N 个但简报不足、设计缺失"
+    kept: List[Dict[str, Any]] = []
+    for r in out:
+        lvl = r.get("risk_level", "safe")
+        if lvl == "blocked":
+            continue
+        kept.append(r)
+
+    stats = dict(state.get("stats") or {})
+    stats["screen"] = {"screened": len(out), "kept": len(kept)}
+    return {"screened": kept, "stats": stats}