Files

142 lines
5.2 KiB
Go

// Copyright 2022 E99p1ant. All rights reserved.
// Use of this source code is governed by a MIT-style
// license that can be found in the LICENSE file.
package censor
import (
"context"
"encoding/json"
"fmt"
"strings"
)
type ForbiddenType string
const (
ForbiddenTypeSpam ForbiddenType = "spam"
ForbiddenTypeAd ForbiddenType = "ad"
ForbiddenTypePolitics ForbiddenType = "politics"
ForbiddenTypeTerrorism ForbiddenType = "terrorism"
ForbiddenTypeAbuse ForbiddenType = "abuse"
ForbiddenTypePorn ForbiddenType = "porn"
ForbiddenTypeFlood ForbiddenType = "flood"
ForbiddenTypeContraband ForbiddenType = "contraband"
ForbiddenTypeMeaningless ForbiddenType = "meaningless"
ForbiddenTypeHarmful ForbiddenType = "harmful"
// ForbiddenTypeCustomized 命中服务侧自定义词库(阿里云控制台配置,七牛无服务端词库)。
// 属运营方显式规则,在 policy.go 中归入底线集合,任何档位都不放行。
ForbiddenTypeCustomized ForbiddenType = "customized"
)
func (f ForbiddenType) String() string {
return map[ForbiddenType]string{
ForbiddenTypeSpam: "含垃圾信息",
ForbiddenTypeAd: "广告",
ForbiddenTypePolitics: "涉政",
ForbiddenTypeTerrorism: "暴恐",
ForbiddenTypeAbuse: "辱骂",
ForbiddenTypePorn: "色情",
ForbiddenTypeFlood: "灌水",
ForbiddenTypeContraband: "违禁",
ForbiddenTypeMeaningless: "无意义",
ForbiddenTypeHarmful: "不良场景",
ForbiddenTypeCustomized: "自定义词库",
}[f]
}
// LabelScore 单个维度(label)的判定与置信度。过去只取首个非 normal 维度,
// 丢掉了其余维度信息,无法按维度差异化处置(例如只拦高置信度的辱骂)。
//
// ⚠️ Score 统一归一化到 0~1:七牛本身即 0~1;阿里云 2.0 无逐标签置信度,只给整条内容的
// riskLevel,由 AliyunTextCensorParser 折算(high=1 / medium=0.6 / low=0.3)后该响应内所有
// 标签共用——有损近似。policy.go 的阈值因此对两家都成立(对阿里云语义为「按风险级别分级」)。
type LabelScore struct {
Label ForbiddenType `json:"label"`
Score float64 `json:"score"`
}
type TextCensorResponse struct {
SourceName string `json:"source_name"`
Pass bool `json:"pass"`
ForbiddenType ForbiddenType `json:"forbidden_type"`
Hint string `json:"hint"`
Confidence float64 `json:"confidence"`
RawResponse json.RawMessage `json:"raw_response"`
Details []LabelScore `json:"details,omitempty"` // 全部非 normal 维度,供严格度策略裁决;由各 parser 从 RawResponse 重建,缓存回放同样有值(故改档位对历史缓存也生效)
// ProviderSuggestion 服务方管控建议(block / review / pass)。不携带维度、无法做
// 置信度收敛,因此只在严格档被采信(见 policy.go)。
ProviderSuggestion string `json:"provider_suggestion,omitempty"`
}
func (r *TextCensorResponse) ToJSON() []byte {
jsonBytes, _ := json.Marshal(r)
return jsonBytes
}
// ErrorMessageFor 按真正命中的维度生成拒绝提示——已移除的 ErrorMessage 固定用首个非
// normal 维度,而触发拦截的可能是另一维度(策略层按档位裁决),会导致文案对不上。
// detail 为 nil(由服务方整体 block 建议触发)时退化为不带维度的基础文案。
func (r *TextCensorResponse) ErrorMessageFor(detail *LabelScore) string {
errorMessage := "内容安全检查不通过"
if detail != nil && detail.Label != "" {
errorMessage += fmt.Sprintf(" [%s]", detail.Label.String())
}
if r.Hint != "" {
errorMessage += fmt.Sprintf(",相关标签:%q", r.Hint)
}
return errorMessage
}
type TextCensor interface {
Censor(ctx context.Context, text string) (*TextCensorResponse, error)
String() string
}
var TrustedHosts = []string{
"xhslink.com", // 小红书
"weibo.com", // 微博
"b23.tv", // Bilibili
"acfun.cn",
"aixifan.com",
"github.com",
// 旧域 tama.guru 保留至 2026-12 到期,过渡期内旧分享链接仍需视为可信。
"tama.guru",
"shiroko.one",
"337845818.best",
"x.com",
"bilibili.com",
}
// RemoveTrustedURL 移除可信域名以免链接被判为广告。
// 仅当可信域名后紧跟路径分隔符(/ ? #)或位于串尾时才视为可信,
// 避免 github.com.evil.tld 被朴素替换误判为 github.com 子域而漏审。
func RemoveTrustedURL(str string) string {
for _, host := range TrustedHosts {
for _, protocol := range []string{"http://", "https://"} {
needle := protocol + host
var b strings.Builder
rest := str
for {
idx := strings.Index(rest, needle)
if idx < 0 {
b.WriteString(rest)
break
}
b.WriteString(rest[:idx])
after := rest[idx+len(needle):]
if len(after) == 0 || after[0] == '/' || after[0] == '?' || after[0] == '#' {
// 可信域名边界:跳过该前缀,从域名之后继续(域名本身被移除,不进入审核)。
rest = after
} else {
// 非可信域名边界(如 github.com.evil):保留前缀,从其后继续扫描。
b.WriteString(needle)
rest = after
}
}
str = b.String()
}
}
return str
}