zuhe guize gai wei tong ju pipei; fangxing duanyu he changjing kaiguan ke zai shezhi ye peizhi

services/content_filter.py

@@ -175,6 +175,13 @@ _WEAK_COMBINATION_TERMS: frozenset[str] = frozenset({"骨骼", "血色"})
 # 送审时，两个词相隔很远并不构成组合风险，默认只在 120 字窗口内才算命中；0 = 不限制。
 _DEFAULT_COMBINATION_WINDOW = 120
 
+# 组合规则还要求左右词「在同一句话里」。列表式提示词（尤其几十条负面提示词）每行
+# 描述一个独立缺陷，「服装撕裂破损」和「身体拉伸扭曲」分属两条、语义无关，却会落进
+# 同一个 120 字窗口被误配（2026-10-17 实例：整段被判「写实血腥」）。
+# 句边界只认换行与句末标点；逗号 / 顿号不算，否则「血肉模糊的人体，写实风格」这种
+# 同一口气写完的真违规也会被切开。可用 local_review.combination_sentence_aware 关闭。
+_SENTENCE_SPLIT = re.compile(r"[\n\r。！？；!?;]+")
+
 # --- 反混淆 ----------------------------------------------------------------
 # 稀有 CJK 区（扩展 A 区、扩展 B 及以后）几乎不出现在正常提示词里，却是
 # 「改革㐀历程」这类插入生僻字绕过词表匹配的惯用手法，规范化时直接剔除。
@@ -205,6 +212,8 @@ class _ReviewSettings(NamedTuple):
     negation_aware: bool
     multilingual_aware: bool
     deobfuscate: bool
+    sentence_aware: bool
+    benign_phrases: frozenset[str]
 
 
 class _SearchText(NamedTuple):
@@ -329,6 +338,47 @@ def _normalize_sensitive_text(value: object) -> str:
     return _canonical_forms(str(value or ""), True)[1]
 
 
+def _sentence_variants(raw: str, *, deobfuscate: bool = True) -> tuple[tuple[_SearchText, ...], ...]:
+    """按句切开分别构造搜索空间，供组合规则做「同句才配对」的近邻判断。
+
+    只在确实切成多句时返回内容；单句返回空元组，调用方回落到整段搜索空间
+    （结果等价，还省一次规范化）。
+    """
+    parts = [part for part in _SENTENCE_SPLIT.split(raw) if part.strip()]
+    if len(parts) <= 1:
+        return ()
+    return tuple((_SearchText(*_canonical_forms(part, deobfuscate)),) for part in parts)
+
+
+def _apply_benign_phrases(
+    variants: tuple[_SearchText, ...],
+    phrases: frozenset[str],
+) -> tuple[_SearchText, ...]:
+    """把配置里的良性短语从搜索空间里抹掉，作用等同扩写内置 `_BENIGN_COMPOUNDS`。
+
+    内置表按「敏感词 → 良性复合词」分组，只能覆盖开发时预想到的搭配。这里是一份
+    平铺的短语清单，站长在设置页加一条就能放行一类误伤，不必改代码重新部署
+    （2026-10-17 用户要求：不要每次误伤都改代码推代码）。
+
+    每个形态各用自己那套归一化结果去替换，所以「服 装 撕 裂」这类插空写法同样有效。
+    """
+    if not phrases:
+        return variants
+    needles = [_canonical_forms(phrase, True) for phrase in phrases]
+    if not needles:
+        return variants
+    masked: list[_SearchText] = []
+    for search in variants:
+        parts = list(search)
+        for forms in needles:
+            for index in range(len(parts)):
+                needle = forms[index]
+                if needle and needle in parts[index]:
+                    parts[index] = parts[index].replace(needle, _BENIGN_COMPOUND_FILLER * len(needle))
+        masked.append(_SearchText(*parts))
+    return tuple(masked)
+
+
 def _search_variants(raw: str, *, deobfuscate: bool = True) -> tuple[_SearchText, ...]:
     """一个请求可能被写成多种等价形态，全部作为搜索空间。"""
     if not raw:
@@ -542,6 +592,17 @@ def _local_review_settings(settings_override: dict[str, object] | None = None) -
     negation_aware = _flag(settings, "negation_aware")
     multilingual_aware = _flag(settings, "multilingual_aware")
     deobfuscate = _flag(settings, "deobfuscate")
+    sentence_aware = _flag(settings, "combination_sentence_aware")
+    raw_phrases = settings.get("benign_phrases")
+    benign_phrases = (
+        frozenset(
+            normalized
+            for phrase in raw_phrases
+            if (normalized := _normalize_sensitive_text(phrase))
+        )
+        if isinstance(raw_phrases, list)
+        else frozenset()
+    )
     combination_window = _coerce_window(settings.get("combination_window", _DEFAULT_COMBINATION_WINDOW))
     raw_strong_terms = settings.get("strong_short_terms")
     strong_terms = (
@@ -564,6 +625,8 @@ def _local_review_settings(settings_override: dict[str, object] | None = None) -
             negation_aware,
             multilingual_aware,
             deobfuscate,
+            sentence_aware,
+            benign_phrases,
         )
     category_terms: dict[str, tuple[str, ...]] = {}
     category_block_terms: set[str] = set()
@@ -614,6 +677,8 @@ def _local_review_settings(settings_override: dict[str, object] | None = None) -
         negation_aware,
         multilingual_aware,
         deobfuscate,
+        sentence_aware,
+        benign_phrases,
     )
 
 
@@ -625,6 +690,7 @@ def _local_rule_match(text: str, words: list[str], settings: dict[str, object] |
     variants = _search_variants(text, deobfuscate=review.deobfuscate)
     if not variants:
         return None
+    variants = _apply_benign_phrases(variants, review.benign_phrases)
     category_match = _first_live_term(variants, review.category_block_terms, negation_aware=negation_aware)
     if category_match:
         return {"rule_id": "category-direct-block", "matched_terms": [category_match]}
@@ -642,16 +708,23 @@ def _local_rule_match(text: str, words: list[str], settings: dict[str, object] |
             continue
         if _configured_word_match(variants, word, negation_aware=negation_aware):
             return {"rule_id": "configured-block-word", "matched_terms": [str(word).strip()]}
-    for rule_id, left_terms, right_terms in review.combination_rules:
-        pair = _closest_rule_pair(
-            variants,
-            left_terms,
-            right_terms,
-            review.combination_window,
-            negation_aware=negation_aware,
-        )
-        if pair:
-            return {"rule_id": rule_id, "matched_terms": [pair[0], pair[1]]}
+    # 组合规则要求左右词在同一句话里。列表式提示词每行描述一个独立缺陷，跨行的两个词
+    # 语义无关却会落进 120 字窗口被误配，因此逐句判断而不是整段扫。
+    search_spaces = (
+        _sentence_variants(text, deobfuscate=review.deobfuscate) if review.sentence_aware else ()
+    ) or (variants,)
+    for search_space in search_spaces:
+        search_space = _apply_benign_phrases(search_space, review.benign_phrases)
+        for rule_id, left_terms, right_terms in review.combination_rules:
+            pair = _closest_rule_pair(
+                search_space,
+                left_terms,
+                right_terms,
+                review.combination_window,
+                negation_aware=negation_aware,
+            )
+            if pair:
+                return {"rule_id": rule_id, "matched_terms": [pair[0], pair[1]]}
     return None
 
 

test/test_content_filter.py

@@ -344,6 +344,46 @@ class ContentFilterTests(unittest.TestCase):
             with self.assertRaises(HTTPException):
                 content_filter.check_request("身体被撕裂的人形，真实画面")
 
+    def test_combination_rule_does_not_pair_terms_across_lines(self) -> None:
+        # 结构性保障，不针对具体词：列表式提示词每行描述一个独立缺陷，
+        # 「身体」在第 3 行、「溃烂」在第 4 行，整段送审时二者落进同一个
+        # 120 字窗口，逐行判断后不再配对。（溃烂没有任何良性复合词豁免，
+        # 所以这条测试验证的是句子边界本身。）
+        with self._patched(self._gore_rule()):
+            content_filter.check_request(
+                "帮我生成一个20岁男性，CG质感，近景构图，浅景深\n"
+                "负面提示词\n"
+                "负面：身体拉伸扭曲、躯干变形\n"
+                "局部溃烂的画面、材质破损\n"
+                "衣物褶皱错乱、布料纹理扭曲\n"
+            )
+            # 同一行里同时出现时，仍然按写实血腥拦截。
+            with self.assertRaises(HTTPException) as raised:
+                content_filter.check_request("身体大面积溃烂，写实画面")
+            self.assertEqual(raised.exception.detail["rule_id"], "写实血腥")
+
+    def test_combination_sentence_aware_can_be_disabled(self) -> None:
+        # 关掉句子边界后回到旧的「整段 120 字窗口」行为，交给用户按需选择。
+        settings = self._gore_rule(combination_sentence_aware=False)
+        with self._patched(settings):
+            with self.assertRaises(HTTPException):
+                content_filter.check_request("身体拉伸扭曲\n局部溃烂的画面")
+
+    def test_configured_benign_phrases_excuse_single_line_lists(self) -> None:
+        # 单行逗号写的负面词列表整段算一句，句子边界救不了，只能靠设置页里的
+        # 「良性短语」豁免：加一条就放行，不需要改代码重新部署。
+        # 用「溃烂」是因为它没有任何内置良性复合词，能干净地验证这条通路。
+        text = "帮我生成一个20岁男性，CG质感\n负面提示词：身体拉伸扭曲、局部溃烂的画面、衣物褶皱错乱"
+        with self._patched(self._gore_rule()):
+            with self.assertRaises(HTTPException):
+                content_filter.check_request(text)
+        with self._patched(self._gore_rule(benign_phrases=["局部溃烂"])):
+            content_filter.check_request(text)
+            # 豁免只针对那一个短语，真血腥语境不受影响。
+            with self.assertRaises(HTTPException):
+                content_filter.check_request("身体大面积溃烂，写实画面")
+
+
     def test_interior_design_prompt_is_not_treated_as_intimate_photography(self) -> None:
         settings = {
             "enabled": True,

web/src/app/settings/components/config-card.tsx

@@ -972,6 +972,35 @@ export function ConfigCard() {
                   />
                   <span className="text-xs text-stone-500">左右两侧命中词必须落在同一段上下文里；0 表示不限距离，默认 120 字。</span>
                 </div>
+                <label className="flex items-start gap-3 text-sm font-medium text-stone-900">
+                  <Checkbox
+                    className="mt-0.5"
+                    checked={config?.local_review?.combination_sentence_aware !== false}
+                    onCheckedChange={(checked) => setLocalReviewField("combination_sentence_aware", Boolean(checked))}
+                    disabled={config?.local_review?.enabled === false}
+                  />
+                  <span>
+                    组合词只在同一句内配对
+                    <span className="mt-1 block text-xs font-normal leading-6 text-stone-500">
+                      负面提示词常一行写一个缺陷，「服装撕裂破损」和「身体拉伸扭曲」分属两条、语义无关，却会落进同一个间距窗口被误判成血腥。开启后只认同句（换行 / 句号 / 分号分隔）内的配对；逗号、顿号连写的长句不受影响，同一句里的真违规照旧拦截。
+                    </span>
+                  </span>
+                </label>
+                <div className="space-y-2">
+                  <label className="text-sm text-stone-700" htmlFor="local-review-benign-phrases">自定义放行短语</label>
+                  <Textarea
+                    id="local-review-benign-phrases"
+                    rows={4}
+                    value={(config?.local_review?.benign_phrases || []).join("\n")}
+                    onChange={(event) => setLocalReviewField("benign_phrases", event.target.value.split("\n").map((phrase) => phrase.trim()).filter(Boolean))}
+                    className="rounded-lg text-sm"
+                    placeholder={"一行一条，例如：\n服装撕裂\n布料撕裂\n唇形崩坏"}
+                    disabled={config?.local_review?.enabled === false}
+                  />
+                  <p className="text-xs text-stone-500">
+                    写在这里的短语会在送审前被整段抹掉，其中的敏感词不再触发拦截。用来放行「服装撕裂」「唇形崩坏」这类在正常提示词里必然出现、但被词表当成风险的多义词搭配。<span className="text-stone-700">保存后立刻生效，不需要重新部署。</span>先用下方「规则测试」粘一条被拦的提示词确认效果。
+                  </p>
+                </div>
               </div>
               <div className="space-y-3 border-t border-stone-200 pt-4">
                 <div>

web/src/app/settings/store.ts

@@ -173,6 +173,8 @@ const DEFAULT_LOCAL_REVIEW = {
   ],
   negation_aware: true,
   combination_window: 120,
+  combination_sentence_aware: true,
+  benign_phrases: [],
   multilingual_aware: true,
   deobfuscate: true,
 };
@@ -224,6 +226,8 @@ function normalizeLocalReview(value: SettingsConfig["local_review"], legacyWords
     })) : DEFAULT_LOCAL_REVIEW.combination_rules,
     negation_aware: value?.negation_aware !== false,
     combination_window: Number.isFinite(rawWindow) && rawWindow >= 0 ? Math.round(rawWindow) : DEFAULT_LOCAL_REVIEW.combination_window,
+    combination_sentence_aware: value?.combination_sentence_aware !== false,
+    benign_phrases: Array.isArray(value?.benign_phrases) ? value.benign_phrases.map((phrase) => String(phrase || "").trim()).filter(Boolean) : [],
     multilingual_aware: value?.multilingual_aware !== false,
     deobfuscate: value?.deobfuscate !== false,
   };
@@ -715,7 +719,7 @@ type SettingsStore = {
   setReversePromptField: (key: "provider" | "base_url" | "api_key" | "model", value: string) => void;
   setProductVisualAgentField: (key: "provider" | "base_url" | "api_key" | "model", value: string) => void;
   setSensitiveWordsText: (value: string) => void;
-  setLocalReviewField: (key: "enabled" | "strong_short_terms" | "categories" | "combination_rules" | "negation_aware" | "combination_window" | "multilingual_aware" | "deobfuscate", value: boolean | number | string[] | NonNullable<SettingsConfig["local_review"]>["categories"] | NonNullable<SettingsConfig["local_review"]>["combination_rules"]) => void;
+  setLocalReviewField: (key: "enabled" | "strong_short_terms" | "categories" | "combination_rules" | "negation_aware" | "combination_window" | "combination_sentence_aware" | "benign_phrases" | "multilingual_aware" | "deobfuscate", value: boolean | number | string[] | NonNullable<SettingsConfig["local_review"]>["categories"] | NonNullable<SettingsConfig["local_review"]>["combination_rules"]) => void;
   setAIReviewField: (key: "enabled" | "base_url" | "api_key" | "model" | "prompt" | "fail_open", value: string | boolean) => void;
   setImageStorageField: (key: keyof ImageStorageSettings, value: string | boolean) => void;
   setProxyRuntimeField: <K extends keyof ProxyRuntimeSettings>(key: K, value: ProxyRuntimeSettings[K]) => void;

web/src/lib/api.ts

@@ -727,6 +727,10 @@ export type SettingsConfig = {
     negation_aware?: boolean;
     /** 组合规则左右两侧命中词的最大间距（按归一化后字符计）。0 = 不限制。默认 120。 */
     combination_window?: number;
+    /** 组合规则要求左右词落在同一句内（换行 / 句号 / 分号分隔）。默认开启。 */
+    combination_sentence_aware?: boolean;
+    /** 送审前整段抹掉的良性短语，用于放行「服装撕裂」这类多数场合无害的搭配。改完即生效。 */
+    benign_phrases?: string[];
     /** 内置跨语言违规词（英/日/韩/俄/德/法/西/越/泰/阿拉伯…）恒先生效。默认开启。 */
     multilingual_aware?: boolean;
     /** 送审前先还原混淆写法：繁→简、URL/HTML/\\u 转义、Base64、稀字剔除、倒序词。默认开启。 */