利用 AI 大模型和自动化工作流,根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.
Python ai-video-generatorcontent-creationffmpeginstagram-reelsllmpythonshort-videosubtitlestext-to-speechtiktokvideo-automationvideo-workflowworkflow-automationyoutube-shorts
Add multilingual social metadata API
harry0703c8d6d769ffdc
diff --git a/app/controllers/v1/llm.py b/app/controllers/v1/llm.pyindex 177dde88..41f6cfb0 100644--- a/app/controllers/v1/llm.py+++ b/app/controllers/v1/llm.py@@ -4,6 +4,8 @@ from app.controllers.v1.base import new_routerfrom app.models.schema import (VideoScriptRequest,VideoScriptResponse,+ VideoSocialMetadataRequest,+ VideoSocialMetadataResponse,VideoTermsRequest,VideoTermsResponse,)@@ -45,3 +47,20 @@ def generate_video_terms(request: Request, body: VideoTermsRequest):)response = {"video_terms": video_terms}return utils.get_response(200, response)+++ "/social-metadata",+ response_model=VideoSocialMetadataResponse,+ summary="Generate social publishing metadata",+)+def generate_video_social_metadata(+ request: Request, body: VideoSocialMetadataRequest+):+ metadata = llm.generate_social_metadata(+ video_subject=body.video_subject,+ video_script=body.video_script,+ language=body.language,+ platform=body.platform,+ )+ return utils.get_response(200, metadata)diff --git a/app/models/schema.py b/app/models/schema.pyindex a57c7f92..eee96618 100644--- a/app/models/schema.py+++ b/app/models/schema.py@@ -178,6 +178,22 @@ class VideoTermsParams:amount: Optional[int] = 5+class VideoSocialMetadataParams:+ """+ {+ "video_subject": "A day in Shanghai",+ "video_script": "",+ "language": "auto",+ "platform": "tiktok"+ }+ """++ video_subject: Optional[str] = Field(default="A day in Shanghai", max_length=500)+ video_script: Optional[str] = Field(default="", max_length=8000)+ language: Optional[str] = Field(default="auto", max_length=64)+ platform: Optional[str] = Field(default="tiktok", max_length=64)++class BaseResponse(BaseModel):status: int = 200message: Optional[str] = "success"@@ -200,6 +216,10 @@ class VideoTermsRequest(VideoTermsParams, BaseModel):pass+class VideoSocialMetadataRequest(VideoSocialMetadataParams, BaseModel):+ pass++##################################################################################################################################################################################################################################################################################################################@@ -284,6 +304,21 @@ class VideoTermsResponse(BaseResponse):}+class VideoSocialMetadataResponse(BaseResponse):+ class Config:+ json_schema_extra = {+ "example": {+ "status": 200,+ "message": "success",+ "data": {+ "title": "A Day in Shanghai You Should Not Miss",+ "caption": "Save this quick Shanghai inspiration and follow for more short travel ideas.",+ "hashtags": ["#shorts", "#travel", "#shanghai", "#viral", "#fyp"],+ },+ },+ }++class BgmRetrieveResponse(BaseResponse):class Config:json_schema_extra = {diff --git a/app/services/llm.py b/app/services/llm.pyindex cbb844ff..2494fd1a 100644--- a/app/services/llm.py+++ b/app/services/llm.py@@ -748,6 +748,272 @@ Please note that you must use English for generating video search terms; Chinesereturn search_terms+# =============================================================================+# Social publishing metadata+#+# 根据视频主题和脚本生成发布到短视频平台时常用的 title、caption 和 hashtags。+# 这块能力只复用现有 LLM provider,不接入任何外部发布服务,也不影响视频生成主链路。+# =============================================================================++# 不同平台的文案长度和 hashtag 数量偏好不同。这里使用保守上限,避免模型返回+# 过长内容后调用方还需要二次裁剪。+SOCIAL_PLATFORMS = {+ "tiktok": {"title_max": 100, "caption_max": 2200, "hashtag_count": 5},+ "youtube_shorts": {"title_max": 100, "caption_max": 5000, "hashtag_count": 3},+ "instagram_reels": {"title_max": 125, "caption_max": 2200, "hashtag_count": 8},+ "facebook_reels": {"title_max": 125, "caption_max": 2200, "hashtag_count": 5},+}+DEFAULT_SOCIAL_PLATFORM = "tiktok"+DEFAULT_SOCIAL_LANGUAGE = "auto"+MAX_SOCIAL_SUBJECT_LENGTH = 500+MAX_SOCIAL_SCRIPT_LENGTH = 8000+MAX_SOCIAL_LANGUAGE_LENGTH = 64++SOCIAL_PLATFORM_LABELS = {+ "tiktok": "TikTok",+ "youtube_shorts": "YouTube Shorts",+ "instagram_reels": "Instagram Reels",+ "facebook_reels": "Facebook Reels",+}++# LLM 不可用时的通用兜底标签。这里故意不绑定某个国家或语种,保证 API+# 对中文、英文、越南语等不同场景都能返回可用结构。+DEFAULT_SOCIAL_HASHTAGS = [+ "#shorts",+ "#viral",+ "#trending",+ "#fyp",+ "#video",+ "#reels",+ "#creator",+ "#content",+]+++def _resolve_social_platform(platform: str | None) -> str:+ value = (platform or "").strip().lower()+ return value if value in SOCIAL_PLATFORMS else DEFAULT_SOCIAL_PLATFORM+++def _normalize_social_language(language: str | None) -> str:+ value = (language or DEFAULT_SOCIAL_LANGUAGE).strip()+ if len(value) > MAX_SOCIAL_LANGUAGE_LENGTH:+ logger.warning(+ "social metadata language is too long and will be truncated to "+ f"{MAX_SOCIAL_LANGUAGE_LENGTH} characters."+ )+ value = value[:MAX_SOCIAL_LANGUAGE_LENGTH]+ return value or DEFAULT_SOCIAL_LANGUAGE+++def _limit_social_text(text: str | None, max_length: int, field_name: str) -> str:+ value = (text or "").strip()+ if len(value) <= max_length:+ return value++ # API 层会限制长度;这里继续兜底,是为了保护内部调用或未来 WebUI+ # 直接调用时不会把超长内容发送给模型,避免 token 成本异常。+ logger.warning(+ f"{field_name} is too long and will be truncated to {max_length} characters."+ )+ return value[:max_length]+++def _social_language_instruction(language: str | None) -> str:+ language = _normalize_social_language(language)+ if language.lower() == DEFAULT_SOCIAL_LANGUAGE:+ return (+ "Use the same language as the video subject and script. If the subject "+ "and script use different languages, prefer the script language."+ )++ return f'Write "title" and "caption" in this language: {language}.'+++def _clamp_text(text, max_length: int) -> str:+ value = ("" if text is None else str(text)).strip()+ if max_length and len(value) > max_length:+ return value[:max_length].rstrip()+ return value+++def _normalize_hashtags(raw, count: int) -> List[str]:+ """+ 将 LLM 返回的 hashtag 统一整理成 `#tag` 格式。++ LLM 可能返回字符串、数组、带空格的词组、重复标签或包含标点的内容。+ 这里集中清洗,可以让接口响应结构稳定,也避免平台发布时出现空标签、+ 重复标签或不符合常见格式的 hashtag。+ """+ if isinstance(raw, str):+ candidates = re.split(r"[\s,]+", raw)+ elif isinstance(raw, (list, tuple)):+ # 数组里的每一项视为一个完整标签,因此 "du lich" 会变成+ # "#dulich",而不是拆成两个标签。+ candidates = [str(entry) for entry in raw]+ else:+ candidates = []++ seen = set()+ result: List[str] = []+ for item in candidates:+ tag = re.sub(r"[^\w]", "", item, flags=re.UNICODE)+ if not tag:+ continue+ key = tag.lower()+ if key in seen:+ continue+ seen.add(key)+ result.append(f"#{tag}")+ if count and len(result) >= count:+ break+ return result+++def build_social_metadata_prompt(+ video_subject: str,+ video_script: str = "",+ language: str = DEFAULT_SOCIAL_LANGUAGE,+ platform: str = DEFAULT_SOCIAL_PLATFORM,+) -> str:+ video_subject = _limit_social_text(+ video_subject, MAX_SOCIAL_SUBJECT_LENGTH, "video_subject"+ )+ video_script = _limit_social_text(+ video_script, MAX_SOCIAL_SCRIPT_LENGTH, "video_script"+ )+ platform = _resolve_social_platform(platform)+ spec = SOCIAL_PLATFORMS[platform]+ label = SOCIAL_PLATFORM_LABELS.get(platform, platform)+ language_instruction = _social_language_instruction(language)++ prompt = f"""+# Role: Short-Video Social Media Copywriter++## Goal+Write engaging publishing metadata for a short video that will be posted on {label}.++## Constraints+1. Respond ONLY with a single valid minified JSON object. No markdown, no code fences, no commentary.+2. The JSON must contain exactly these keys: "title", "caption", "hashtags".+3. "title": a catchy hook, at most {spec['title_max']} characters.+4. "caption": an engaging description that ends with a call to action, at most {spec['caption_max']} characters. Do not put hashtags inside the caption.+5. "hashtags": a JSON array of exactly {spec['hashtag_count']} strings. Each must start with "#", contain no spaces, and be relevant to the topic and to {label}.+6. {language_instruction}++## Output Example+{{"title":"...","caption":"...","hashtags":["#example","#video"]}}++## Context+### Video Subject+{video_subject}++### Video Script+{video_script}+""".strip()+ return prompt+++def _parse_social_metadata(response: str, platform: str) -> dict:+ spec = SOCIAL_PLATFORMS[_resolve_social_platform(platform)]++ data = None+ try:+ data = json.loads(response)+ except Exception:+ # 部分模型会在 JSON 外层包一段说明文字或 markdown fence。+ # API 调用方只需要稳定结构,所以这里尝试提取第一个 JSON object。+ match = re.search(r"\{.*\}", response or "", re.DOTALL)+ if match:+ data = json.loads(match.group())++ if not isinstance(data, dict):+ raise ValueError("social metadata response is not a JSON object")++ title = _clamp_text(data.get("title", ""), spec["title_max"])+ caption = _clamp_text(data.get("caption", ""), spec["caption_max"])+ hashtags = _normalize_hashtags(data.get("hashtags", []), spec["hashtag_count"])++ if not title and not caption:+ raise ValueError("social metadata response is missing both title and caption")++ return {"title": title, "caption": caption, "hashtags": hashtags}+++def _fallback_social_metadata(+ video_subject: str, video_script: str, platform: str+) -> dict:+ spec = SOCIAL_PLATFORMS[_resolve_social_platform(platform)]+ subject = (video_subject or "").strip()+ script = (video_script or "").strip()++ title = subject+ if not title and script:+ # 没有主题时,用脚本第一句兜底生成 title,避免接口返回空标题。+ title = re.split(r"(?<=[.!?。!?])\s+", script)[0]++ return {+ "title": _clamp_text(title, spec["title_max"]),+ "caption": _clamp_text(script or subject, spec["caption_max"]),+ "hashtags": _normalize_hashtags(+ DEFAULT_SOCIAL_HASHTAGS, spec["hashtag_count"]+ ),+ }+++def generate_social_metadata(+ video_subject: str,+ video_script: str = "",+ language: str = DEFAULT_SOCIAL_LANGUAGE,+ platform: str = DEFAULT_SOCIAL_PLATFORM,+) -> dict:+ """+ 生成短视频发布文案元数据。++ 返回结构固定为 `{"title": str, "caption": str, "hashtags": List[str]}`。+ 如果 LLM 不可用或返回格式异常,会降级为通用启发式结果,保证 API+ 调用方始终拿到可展示、可发布前编辑的数据结构。+ """+ platform = _resolve_social_platform(platform)+ language = _normalize_social_language(language)+ video_subject = _limit_social_text(+ video_subject, MAX_SOCIAL_SUBJECT_LENGTH, "video_subject"+ )+ video_script = _limit_social_text(+ video_script, MAX_SOCIAL_SCRIPT_LENGTH, "video_script"+ )+ prompt = build_social_metadata_prompt(+ video_subject=video_subject,+ video_script=video_script,+ language=language,+ platform=platform,+ )+ logger.info(+ f"generating social metadata: platform={platform}, language={language}"+ )++ response = ""+ for i in range(_max_retries):+ try:+ response = _generate_response(prompt)+ if isinstance(response, str) and "Error: " in response:+ logger.error(f"failed to generate social metadata: {response}")+ break+ metadata = _parse_social_metadata(response, platform)+ logger.success(f"completed: \n{metadata}")+ return metadata+ except Exception as e:+ logger.warning(f"failed to parse social metadata: {str(e)}")++ if i < _max_retries - 1:+ logger.warning(+ f"failed to generate social metadata, trying again... {i + 1}"+ )++ logger.warning("falling back to heuristic social metadata")+ return _fallback_social_metadata(video_subject, video_script, platform)++if __name__ == "__main__":video_subject = "生命的意义是什么"script = generate_script(diff --git a/test/services/test_llm.py b/test/services/test_llm.pyindex fa70aa4b..70d5acda 100644--- a/test/services/test_llm.py+++ b/test/services/test_llm.py@@ -11,7 +11,7 @@ from pydantic import ValidationErrorsys.path.insert(0, str(Path(__file__).parent.parent.parent))from app.config import config-from app.models.schema import VideoScriptRequest+from app.models.schema import VideoScriptRequest, VideoSocialMetadataRequestfrom app.services import llm@@ -614,6 +614,171 @@ class TestRuntimeEnvironmentDetection(unittest.TestCase):self.assertEqual(config.get_container_default_gateway_ip(str(route_path)), "")+class TestSocialMetadata(unittest.TestCase):+ """通用短视频发布文案元数据生成。"""++ def test_build_prompt_auto_language_uses_source_language(self):+ """+ language 默认 auto 时,不应该固定成某个国家或语种,而是让模型+ 跟随视频主题和脚本的语言,扩大 API 适用范围。+ """+ prompt = llm.build_social_metadata_prompt(+ video_subject="上海一日游",+ video_script="今天带你快速看完上海经典路线。",+ language="auto",+ platform="tiktok",+ )++ self.assertIn("TikTok", prompt)+ self.assertIn("Use the same language as the video subject and script", prompt)+ self.assertIn("上海一日游", prompt)+ self.assertIn("array of exactly 5 strings", prompt)++ def test_build_prompt_accepts_explicit_language(self):+ prompt = llm.build_social_metadata_prompt(+ video_subject="Coffee tips",+ language="en-US",+ platform="youtube_shorts",+ )++ self.assertIn("YouTube Shorts", prompt)+ self.assertIn('Write "title" and "caption" in this language: en-US', prompt)+ self.assertIn("array of exactly 3 strings", prompt)++ def test_unknown_platform_falls_back_to_tiktok(self):+ prompt = llm.build_social_metadata_prompt(+ video_subject="x",+ platform="unsupported-platform",+ )++ self.assertIn("TikTok", prompt)++ def test_normalize_hashtags_from_string_dedupes_and_clamps(self):+ tags = llm._normalize_hashtags("#fyp fyp, trending #Trending viral", count=2)++ self.assertEqual(tags, ["#fyp", "#trending"])++ def test_normalize_hashtags_from_list_keeps_unicode_letters(self):+ tags = llm._normalize_hashtags(+ ["上海 旅行", "#việt nam", " ", "@bad!chars"], count=5+ )++ self.assertEqual(tags, ["#上海旅行", "#việtnam", "#badchars"])++ def test_parse_social_metadata_recovers_embedded_json(self):+ raw = 'Sure: {"title":"T","caption":"C","hashtags":["#x"]} thanks'+ result = llm._parse_social_metadata(raw, "tiktok")++ self.assertEqual(result["title"], "T")+ self.assertEqual(result["caption"], "C")+ self.assertEqual(result["hashtags"], ["#x"])++ def test_parse_social_metadata_requires_title_or_caption(self):+ with self.assertRaises(ValueError):+ llm._parse_social_metadata('{"hashtags":["#x"]}', "tiktok")++ def test_generate_social_metadata_uses_llm_response(self):+ payload = (+ '{"title":"上海一日游","caption":"收藏这条路线,下次直接出发!",'+ '"hashtags":["#上海","#旅行","#shorts"]}'+ )+ with patch.object(llm, "_generate_response", return_value=payload):+ result = llm.generate_social_metadata(+ video_subject="上海一日游",+ video_script="今天带你快速看完上海经典路线。",+ language="zh-CN",+ platform="tiktok",+ )++ self.assertEqual(result["title"], "上海一日游")+ self.assertEqual(result["caption"], "收藏这条路线,下次直接出发!")+ self.assertEqual(result["hashtags"], ["#上海", "#旅行", "#shorts"])++ def test_generate_social_metadata_falls_back_to_generic_hashtags(self):+ with patch.object(+ llm, "_generate_response", return_value="Error: api_key is not set"+ ):+ result = llm.generate_social_metadata(+ video_subject="Coffee tips",+ video_script="Save these three coffee tips.",+ platform="instagram_reels",+ )++ self.assertEqual(result["title"], "Coffee tips")+ self.assertEqual(result["caption"], "Save these three coffee tips.")+ self.assertEqual(len(result["hashtags"]), 8)+ self.assertEqual(result["hashtags"][0], "#shorts")++ def test_request_model_defaults_to_auto_language_tiktok(self):+ body = VideoSocialMetadataRequest(video_subject="Test")++ self.assertEqual(body.language, "auto")+ self.assertEqual(body.platform, "tiktok")++ def test_request_model_rejects_oversized_social_metadata_fields(self):+ """+ 外部 API 不能接受无限长的脚本和语言参数,否则会直接放大 LLM+ token 成本。schema 层先拦截,服务层再做内部调用兜底。+ """+ with self.assertRaises(ValidationError):+ VideoSocialMetadataRequest(video_subject="x" * 501)++ with self.assertRaises(ValidationError):+ VideoSocialMetadataRequest(video_subject="x", video_script="x" * 8001)++ with self.assertRaises(ValidationError):+ VideoSocialMetadataRequest(video_subject="x", language="x" * 65)++ def test_build_prompt_clamps_direct_service_inputs(self):+ prompt = llm.build_social_metadata_prompt(+ video_subject="x" * 600,+ video_script="y" * 9000,+ language="en",+ )++ self.assertIn("x" * llm.MAX_SOCIAL_SUBJECT_LENGTH, prompt)+ self.assertNotIn("x" * (llm.MAX_SOCIAL_SUBJECT_LENGTH + 1), prompt)+ self.assertIn("y" * llm.MAX_SOCIAL_SCRIPT_LENGTH, prompt)+ self.assertNotIn("y" * (llm.MAX_SOCIAL_SCRIPT_LENGTH + 1), prompt)++ def test_social_metadata_endpoint_response_shape(self):+ from fastapi.testclient import TestClient++ from app.asgi import app++ request_body = {+ "video_subject": "Tokyo coffee shops",+ "video_script": "Three quiet coffee shops for your next Tokyo morning.",+ "language": "en",+ "platform": "youtube_shorts",+ }+ llm_response = (+ '{"title":"3 Quiet Tokyo Coffee Shops",'+ '"caption":"Save these spots for your next Tokyo morning.",'+ '"hashtags":["#Tokyo","#Coffee","#Shorts"]}'+ )++ with patch.object(llm, "_generate_response", return_value=llm_response):+ response = TestClient(app).post(+ "/api/v1/social-metadata",+ json=request_body,+ )++ self.assertEqual(response.status_code, 200)+ self.assertEqual(+ response.json(),+ {+ "status": 200,+ "message": "success",+ "data": {+ "title": "3 Quiet Tokyo Coffee Shops",+ "caption": "Save these spots for your next Tokyo morning.",+ "hashtags": ["#Tokyo", "#Coffee", "#Shorts"],+ },+ },+ )++FOUNDRY_KEY = os.environ.get("ANTHROPIC_FOUNDRY_API_KEY", "")FOUNDRY_BASE = "https://amanrai-test-resource.services.ai.azure.com/anthropic"FOUNDRY_MODEL = "azure_ai/claude-sonnet-4-6"