|
|
@@ -0,0 +1,274 @@
|
|
|
+import os
|
|
|
+import tempfile
|
|
|
+from urllib.parse import urlparse
|
|
|
+
|
|
|
+import requests
|
|
|
+from fastapi import APIRouter, HTTPException
|
|
|
+
|
|
|
+from ..core.ark_client import get_client
|
|
|
+from ..config.config import Config
|
|
|
+from ..schemas.chat import ModerationRequest, ModerationResponse
|
|
|
+from ..db.mongo import save_chat_log
|
|
|
+
|
|
|
+router = APIRouter()
|
|
|
+
|
|
|
+# 待审核文本最大长度限制(超出则拒绝,避免超长文本拖慢模型响应)
|
|
|
+MAX_CONTENT_LENGTH = 20000
|
|
|
+# 远程媒体下载大小限制(火山方舟视频上限 50MB)
|
|
|
+MAX_VIDEO_SIZE = 50 * 1024 * 1024
|
|
|
+MAX_IMAGE_SIZE = 20 * 1024 * 1024
|
|
|
+# 下载超时(连接超时, 读取超时)
|
|
|
+_DOWNLOAD_TIMEOUT = (10, 60)
|
|
|
+
|
|
|
+# 内容审核系统提示词:约束模型仅输出 true / false,降低解析歧义并提升响应速度
|
|
|
+_MODERATION_PROMPT = """你是一个严格、专业的内容安全审核员。请判断用户提供的内容(文本、图片或视频画面)是否包含违规信息。
|
|
|
+
|
|
|
+违规内容包括以下任一类别:
|
|
|
+1. 涉黄内容:色情、低俗、性暗示、裸露等;
|
|
|
+2. 暴力恐怖内容:血腥、暴力、恐怖主义、极端主义等;
|
|
|
+3. 涉政敏感内容:政治敏感、反动言论、危害国家安全等;
|
|
|
+4. 其他不良信息:违法犯罪、赌博、诈骗、毒品、辱骂攻击、封建迷信等;
|
|
|
+5. 广告推广内容:商业营销、引流推广、联系方式诱导、二维码等。
|
|
|
+
|
|
|
+判断规则:
|
|
|
+- 只要内容命中上述任意一类,即判定为违规;
|
|
|
+- 对图片/视频,需结合画面主体、文字、场景等综合判断;
|
|
|
+- 内容健康、正常、合规的判定为合规;
|
|
|
+- 仅依据所提供的内容判断,不做额外联想或补全。
|
|
|
+
|
|
|
+输出要求(务必严格遵守):
|
|
|
+- 只输出一个单词:true 或 false;
|
|
|
+- true 表示检测到违规内容,false 表示合规内容;
|
|
|
+- 禁止输出任何解释、标点、空格或其他多余字符。"""
|
|
|
+
|
|
|
+
|
|
|
+def _extract_text(response) -> str:
|
|
|
+ """从 responses API 返回结果中提取文本内容"""
|
|
|
+ result = ""
|
|
|
+ for item in response.output:
|
|
|
+ if hasattr(item, 'type') and item.type == 'message' and hasattr(item, 'content'):
|
|
|
+ if isinstance(item.content, list):
|
|
|
+ for content_item in item.content:
|
|
|
+ if hasattr(content_item, 'text'):
|
|
|
+ result += content_item.text
|
|
|
+ else:
|
|
|
+ result += str(item.content)
|
|
|
+ return result
|
|
|
+
|
|
|
+
|
|
|
+def _parse_moderation_result(text: str) -> bool:
|
|
|
+ """将模型输出解析为布尔值:True=违规,False=合规"""
|
|
|
+ cleaned = text.strip().lower()
|
|
|
+ if cleaned.startswith("true"):
|
|
|
+ return True
|
|
|
+ if cleaned.startswith("false"):
|
|
|
+ return False
|
|
|
+ # 容错:模型可能附带少量多余字符,取首次出现的关键词
|
|
|
+ has_true = "true" in cleaned
|
|
|
+ has_false = "false" in cleaned
|
|
|
+ if has_true and not has_false:
|
|
|
+ return True
|
|
|
+ if has_false and not has_true:
|
|
|
+ return False
|
|
|
+ # 无法明确解析时抛错,交由调用方重试,避免误判放行违规内容
|
|
|
+ raise ValueError(f"无法解析审核结果: {text[:50]}")
|
|
|
+
|
|
|
+
|
|
|
+def _resolve_ext(url: str, content: bytes, default: str) -> str:
|
|
|
+ """推断文件扩展名:优先取 URL 路径后缀,其次按文件魔数嗅探,最后用默认值。
|
|
|
+ Files API 依据扩展名识别媒体类型,因此需给出正确后缀。"""
|
|
|
+ ext = os.path.splitext(urlparse(url).path)[1].lower()
|
|
|
+ known = (".mp4", ".mov", ".webm", ".mkv", ".avi", ".flv",
|
|
|
+ ".jpg", ".jpeg", ".png", ".webp", ".gif", ".bmp")
|
|
|
+ if ext in known:
|
|
|
+ return ext
|
|
|
+ # 文件魔数嗅探(URL 无有效后缀时,如 COS 的 ?imageMogr2 处理链接)
|
|
|
+ if content[:3] == b"\xff\xd8\xff":
|
|
|
+ return ".jpg"
|
|
|
+ if content[:8] == b"\x89PNG\r\n\x1a\n":
|
|
|
+ return ".png"
|
|
|
+ if content[:4] == b"RIFF" and content[8:12] == b"WEBP":
|
|
|
+ return ".webp"
|
|
|
+ if content[:6] in (b"GIF87a", b"GIF89a"):
|
|
|
+ return ".gif"
|
|
|
+ if content[4:8] == b"ftyp":
|
|
|
+ return ".mp4"
|
|
|
+ return default
|
|
|
+
|
|
|
+
|
|
|
+def _upload_remote_to_ark(client, url: str, kind: str) -> str:
|
|
|
+ """下载远程 URL 并上传到火山方舟 Files API,返回 file_id。
|
|
|
+
|
|
|
+ 直接把 COS 等第三方 URL 传给模型时,若源站返回的 Content-Type 为
|
|
|
+ application/octet-stream,方舟会因无法识别媒体类型而报 400。通过
|
|
|
+ 下载 + Files API 上传(按扩展名识别类型)可彻底绕开该限制。
|
|
|
+
|
|
|
+ kind: 'video' | 'image',用于大小限制与默认扩展名。
|
|
|
+ """
|
|
|
+ max_size = MAX_VIDEO_SIZE if kind == "video" else MAX_IMAGE_SIZE
|
|
|
+ try:
|
|
|
+ resp = requests.get(url, timeout=_DOWNLOAD_TIMEOUT, stream=True)
|
|
|
+ resp.raise_for_status()
|
|
|
+ content = b""
|
|
|
+ for chunk in resp.iter_content(chunk_size=1 << 20):
|
|
|
+ if not chunk:
|
|
|
+ continue
|
|
|
+ content += chunk
|
|
|
+ if len(content) > max_size:
|
|
|
+ raise HTTPException(
|
|
|
+ status_code=400,
|
|
|
+ detail=f"{kind}文件超过大小限制({max_size // 1024 // 1024}MB)",
|
|
|
+ )
|
|
|
+ except HTTPException:
|
|
|
+ raise
|
|
|
+ except Exception as e:
|
|
|
+ raise HTTPException(status_code=400, detail=f"无法下载{kind}资源: {str(e)}")
|
|
|
+
|
|
|
+ if not content:
|
|
|
+ raise HTTPException(status_code=400, detail=f"{kind}资源内容为空")
|
|
|
+
|
|
|
+ ext = _resolve_ext(url, content, ".mp4" if kind == "video" else ".jpg")
|
|
|
+ filename = f"moderation_{kind}{ext}"
|
|
|
+ tmp_path = None
|
|
|
+ try:
|
|
|
+ with tempfile.NamedTemporaryFile(delete=False, suffix=ext) as tmp:
|
|
|
+ tmp.write(content)
|
|
|
+ tmp_path = tmp.name
|
|
|
+ with open(tmp_path, "rb") as f:
|
|
|
+ uploaded = client.files.create(file=(filename, f), purpose="user_data")
|
|
|
+ return uploaded.id
|
|
|
+ except Exception as e:
|
|
|
+ raise HTTPException(status_code=502, detail=f"上传{kind}到审核服务失败: {str(e)}")
|
|
|
+ finally:
|
|
|
+ if tmp_path and os.path.exists(tmp_path):
|
|
|
+ os.unlink(tmp_path)
|
|
|
+
|
|
|
+
|
|
|
+def _resolve_type(request: ModerationRequest) -> str:
|
|
|
+ """确定审核类型:优先使用显式 type,否则根据已传参数自动推断"""
|
|
|
+ if request.type:
|
|
|
+ t = request.type.strip().lower()
|
|
|
+ if t not in ("text", "image", "video"):
|
|
|
+ raise HTTPException(status_code=400, detail=f"不支持的审核类型: {request.type}")
|
|
|
+ return t
|
|
|
+ if request.video_url:
|
|
|
+ return "video"
|
|
|
+ if request.image_url:
|
|
|
+ return "image"
|
|
|
+ if request.content:
|
|
|
+ return "text"
|
|
|
+ raise HTTPException(status_code=400, detail="请提供待审核的内容(content / image_url / video_url)")
|
|
|
+
|
|
|
+
|
|
|
+def _build_input_content(mod_type: str, request: ModerationRequest, client) -> list:
|
|
|
+ """根据审核类型构造 Responses API 多模态 input content,并做参数校验。
|
|
|
+ 图片/视频统一走 Files API(file_id),避免第三方 URL 的 mimetype 校验问题。"""
|
|
|
+ if mod_type == "text":
|
|
|
+ content = (request.content or "").strip()
|
|
|
+ if not content:
|
|
|
+ raise HTTPException(status_code=400, detail="待审核文本不能为空")
|
|
|
+ if len(content) > MAX_CONTENT_LENGTH:
|
|
|
+ raise HTTPException(
|
|
|
+ status_code=400,
|
|
|
+ detail=f"待审核文本超过长度限制({MAX_CONTENT_LENGTH} 字符),请分段提交",
|
|
|
+ )
|
|
|
+ return [{"type": "input_text", "text": content}]
|
|
|
+
|
|
|
+ if mod_type == "image":
|
|
|
+ image_url = (request.image_url or "").strip()
|
|
|
+ if not image_url:
|
|
|
+ raise HTTPException(status_code=400, detail="图片审核需提供 image_url")
|
|
|
+ file_id = _upload_remote_to_ark(client, image_url, "image")
|
|
|
+ return [
|
|
|
+ {"type": "input_image", "file_id": file_id},
|
|
|
+ {"type": "input_text", "text": "请审核这张图片的画面内容是否违规。"},
|
|
|
+ ]
|
|
|
+
|
|
|
+ # video:对视频画面 + 封面图一起审核
|
|
|
+ video_url = (request.video_url or "").strip()
|
|
|
+ thumb_url = (request.thumb_url or "").strip()
|
|
|
+ if not video_url:
|
|
|
+ raise HTTPException(status_code=400, detail="视频审核需提供 video_url")
|
|
|
+ if not thumb_url:
|
|
|
+ raise HTTPException(status_code=400, detail="视频审核需提供封面图 thumb_url")
|
|
|
+ video_file_id = _upload_remote_to_ark(client, video_url, "video")
|
|
|
+ thumb_file_id = _upload_remote_to_ark(client, thumb_url, "image")
|
|
|
+ return [
|
|
|
+ {"type": "input_video", "file_id": video_file_id},
|
|
|
+ {"type": "input_image", "file_id": thumb_file_id},
|
|
|
+ {"type": "input_text", "text": "请综合审核这段视频的画面及封面图内容是否违规。"},
|
|
|
+ ]
|
|
|
+
|
|
|
+
|
|
|
+def _log_target(mod_type: str, request: ModerationRequest) -> str:
|
|
|
+ """生成审核日志的目标标识(文本截断 / 媒体 URL)"""
|
|
|
+ if mod_type == "text":
|
|
|
+ return (request.content or "")[:500]
|
|
|
+ if mod_type == "image":
|
|
|
+ return f"[image] {request.image_url}"
|
|
|
+ return f"[video] video={request.video_url} thumb={request.thumb_url}"
|
|
|
+
|
|
|
+
|
|
|
+@router.post("/check", response_model=ModerationResponse)
|
|
|
+async def moderate_content(request: ModerationRequest):
|
|
|
+ """
|
|
|
+ 公共内容审核接口(无需登录认证),供外部系统调用。
|
|
|
+ 支持三种审核类型,利用 AI 模型检测是否包含涉黄、暴恐、涉政、其他不良、广告推广等违规内容:
|
|
|
+ - text:审核文本内容(content)
|
|
|
+ - image:审核图片(image_url)
|
|
|
+ - video:审核视频画面 + 封面图(video_url + thumb_url)
|
|
|
+
|
|
|
+ 图片/视频通过 Files API 上传换取 file_id 后送审,规避第三方 URL 的 mimetype 校验问题。
|
|
|
+ 未显式传 type 时,根据已传参数自动推断(video_url > image_url > content)。
|
|
|
+ 返回:sensitive=true 表示检测到违规内容,false 表示合规内容。
|
|
|
+ """
|
|
|
+ mod_type = _resolve_type(request)
|
|
|
+ log_target = _log_target(mod_type, request)
|
|
|
+
|
|
|
+ try:
|
|
|
+ client = get_client()
|
|
|
+ except Exception as e:
|
|
|
+ raise HTTPException(status_code=503, detail=f"审核服务初始化失败: {str(e)}")
|
|
|
+
|
|
|
+ # 图片/视频会在此处下载远程资源并上传 Files API(可能抛 400/502)
|
|
|
+ input_content = _build_input_content(mod_type, request, client)
|
|
|
+
|
|
|
+ system_msg = {"role": "system", "content": [{"type": "input_text", "text": _MODERATION_PROMPT}]}
|
|
|
+ user_msg = {"role": "user", "content": input_content}
|
|
|
+
|
|
|
+ try:
|
|
|
+ response = client.responses.create(
|
|
|
+ model=Config.MODEL_NAME,
|
|
|
+ input=[system_msg, user_msg],
|
|
|
+ stream=False,
|
|
|
+ store=False, # 无状态审核,不保存上下文,提升性能
|
|
|
+ text={"format": {"type": "text"}},
|
|
|
+ )
|
|
|
+ except Exception as e:
|
|
|
+ save_chat_log(
|
|
|
+ user_id="moderation",
|
|
|
+ question=log_target,
|
|
|
+ stream_mode=False,
|
|
|
+ status="error",
|
|
|
+ error=str(e),
|
|
|
+ )
|
|
|
+ raise HTTPException(status_code=500, detail=f"内容审核服务异常: {str(e)}")
|
|
|
+
|
|
|
+ raw_text = _extract_text(response)
|
|
|
+ try:
|
|
|
+ is_sensitive = _parse_moderation_result(raw_text)
|
|
|
+ except ValueError as e:
|
|
|
+ save_chat_log(
|
|
|
+ user_id="moderation",
|
|
|
+ question=log_target,
|
|
|
+ stream_mode=False,
|
|
|
+ raw_response=repr(response),
|
|
|
+ status="error",
|
|
|
+ error=str(e),
|
|
|
+ )
|
|
|
+ raise HTTPException(status_code=500, detail="内容审核结果解析失败,请重试")
|
|
|
+
|
|
|
+ return ModerationResponse(sensitive=is_sensitive)
|
|
|
+
|
|
|
+
|
|
|
+router.tags = ["内容审核"]
|