feat(site-import): 行为骨架模板与 few-shot 语料

- template/ 四件套：system/extractor/evaluator 骨架 + placeholders.json 占位符清单（system 15 个占位符，固定段与 jiancai 原文逐字一致——行级来源审计 0 缺失） - fewshot/ 语料摘录：yotaienergys/televicsic 两生产租户四件套行业化片段（10592 字节 ≤18000 上限），README 含来源声明与数据治理记档 - tests/test_site_import_template.py：行业残留黑名单/占位符对称/红线锚点/ 行级来源/fewshot 尺寸与用途段守卫，7 条 Co-Authored-By: Claude Code <noreply@anthropic.com>
2026-10-20 19:35:08 +08:00
parent 74c2915f25
commit 70e1355489
7 changed files with 483 additions and 0 deletions
@@ -0,0 +1,70 @@
+"""site_import 骨架模板与 few-shot 语料守卫（Task 4）。"""
+import json
+import re
+from pathlib import Path
+
+TPL_DIR = Path(__file__).resolve().parents[1] / "app" / "platform" / "site_import" / "template"
+FS = Path(__file__).resolve().parents[1] / "app" / "platform" / "site_import" / "fewshot" / "fewshot_excerpts.md"
+JIANCAI_SYSTEM = Path(__file__).resolve().parents[1] / "seed_profiles" / "jiancai" / "prompts" / "system.md"
+
+# 行业专属词黑名单（jiancai 管道 / yotai 储能 / televics 炉辊三行业的专属词；
+# WhatsApp、manager、quotation 等通用外贸词不在其列）
+INDUSTRY_WORDS = [
+    "管道", "pipe", "HESUPIPE", "储能", "storage", "炉辊", "roller",
+    "TELEVIC", "YOTAI", "HDPE", "PVC", "PE100", "SDR", "DN110",
+    "给水", "燃气", "排污", "灌溉", "波纹管", "球墨铸铁",
+]
+
+
+class TestTemplatePurity:
+    def test_骨架无行业残留(self):
+        for fname in ("system", "extractor", "evaluator"):
+            t = (TPL_DIR / f"{fname}_template.md").read_text(encoding="utf-8")
+            low = t.lower()
+            for w in INDUSTRY_WORDS:
+                assert w.lower() not in low, f"{fname}_template.md 行业残留: {w}"
+
+    def test_占位符与清单一致(self):
+        ph = json.loads((TPL_DIR / "placeholders.json").read_text(encoding="utf-8"))
+        assert set(ph) == {"system", "extractor", "evaluator"}
+        for fname, keys in ph.items():
+            t = (TPL_DIR / f"{fname}_template.md").read_text(encoding="utf-8")
+            for k in keys:
+                assert "{{" + k + "}}" in t, f"{fname} 缺占位符 {k}"
+        # 反向：模板中每个 {{xxx}} 都在清单
+        for fname, keys in ph.items():
+            t = (TPL_DIR / f"{fname}_template.md").read_text(encoding="utf-8")
+            found = set(re.findall(r"\{\{([^}]+)\}\}", t))
+            assert found == set(keys), f"{fname} 占位符不对称: {found ^ set(keys)}"
+
+    def test_红线段锚点齐全(self):
+        t = (TPL_DIR / "system_template.md").read_text(encoding="utf-8")
+        for anchor in ["语言规则（最高优先级）", "职能边界", "红线（任何情况下不可违反）",
+                       "不能出现任何价格数字", "施压", "禁止说的话", "联系方式引导策略",
+                       "12 hours", "国家代码"]:
+            assert anchor in t, f"缺锚点: {anchor}"
+
+    def test_固定段行级来源(self):
+        """非占位行应逐字存在于 jiancai 原文（允许空行/标题差异 ±5% 行）。"""
+        j = JIANCAI_SYSTEM.read_text(encoding="utf-8").splitlines()
+        jset = set(l.strip() for l in j)
+        t = (TPL_DIR / "system_template.md").read_text(encoding="utf-8").splitlines()
+        missing = [l.strip() for l in t
+                   if l.strip() and "{{" not in l and "#" not in l[:2]
+                   and l.strip() not in jset]
+        # 允许少量结构性改写（拆段标题等）——阈值 ≤8 行，且打印出来人工可核
+        assert len(missing) <= 8, f"来源外行过多: {missing}"
+
+
+class TestFewshot:
+    def test_尺寸上限(self):
+        assert len(FS.read_bytes()) <= 18000
+
+    def test_两租户都在(self):
+        s = FS.read_text(encoding="utf-8")
+        assert "yotaienergys" in s and "televicsic" in s
+
+    def test_用途段落齐(self):
+        s = FS.read_text(encoding="utf-8")
+        for sec in ["身份", "追问", "好差", "事实卡", "提取器", "评估器"]:
+            assert sec in s, f"fewshot 缺用途段: {sec}"