| 1 | import json |
| 2 | import logging |
| 3 | import re |
| 4 | import requests |
| 5 | from typing import List |
| 6 | |
| 7 | from loguru import logger |
| 8 | from openai import AzureOpenAI, OpenAI |
| 9 | from openai.types.chat import ChatCompletion |
| 10 | |
| 11 | from app.config import config |
| 12 | |
| 13 | _max_retries = 5 |
| 14 | _DEFAULT_GEMINI_MODEL = "gemini-2.5-flash" |
| 15 | _DEPRECATED_GEMINI_MODELS = {"gemini-pro", "gemini-1.0-pro"} |
| 16 | MIN_SCRIPT_PARAGRAPH_NUMBER = 1 |
| 17 | MAX_SCRIPT_PARAGRAPH_NUMBER = 10 |
| 18 | MAX_SCRIPT_PROMPT_LENGTH = 2000 |
| 19 | MAX_SCRIPT_SYSTEM_PROMPT_LENGTH = 8000 |
| 20 | _THINK_BLOCK_RE = re.compile(r"<think\b[^>]*>.*?</think>", re.IGNORECASE | re.DOTALL) |
| 21 | _UNCLOSED_THINK_BLOCK_RE = re.compile(r"<think\b[^>]*>.*$", re.IGNORECASE | re.DOTALL) |
| 22 | _URL_USERINFO_RE = re.compile(r"((?:https?|wss?)://)([^/\s?#@]*:[^/\s?#@]*@)", re.IGNORECASE) |
| 23 | _SENSITIVE_QUERY_RE = re.compile( |
| 24 | r"([?&](?:api[_-]?key|access[_-]?token|token|key|secret|password)=)([^&#\s]+)", |
| 25 | re.IGNORECASE, |
| 26 | ) |
| 27 | |
| 28 | DEFAULT_SCRIPT_SYSTEM_PROMPT = """ |
| 29 | # Role: Video Script Generator |
| 30 | |
| 31 | ## Goals: |
| 32 | Generate a script for a video, depending on the subject of the video. |
| 33 | |
| 34 | ## Constrains: |
| 35 | 1. the script is to be returned as a string with the specified number of paragraphs. |
| 36 | 2. do not under any circumstance reference this prompt in your response. |
| 37 | 3. get straight to the point, don't start with unnecessary things like, "welcome to this video". |
| 38 | 4. you must not include any type of markdown or formatting in the script, never use a title. |
| 39 | 5. only return the raw content of the script. |
| 40 | 6. do not include "voiceover", "narrator" or similar indicators of what should be spoken at the beginning of each paragraph or line. |
| 41 | 7. you must not mention the prompt, or anything about the script itself. also, never talk about the amount of paragraphs or lines. just write the script. |
| 42 | 8. respond in the same language as the video subject. |
| 43 | """.strip() |
| 44 | |
| 45 | |
| 46 | def _normalize_text_response(content, llm_provider: str) -> str: |
| 47 | # 不同 LLM SDK 在异常或被拦截场景下，可能返回 None、空字符串， |
| 48 | # 甚至返回非字符串对象。这里统一做兜底校验，避免后续直接调用 |
| 49 | # `.replace()` 时抛出 `NoneType` 之类的属性错误。 |
| 50 | if content is None: |
| 51 | raise ValueError(f"[{llm_provider}] returned empty text content") |
| 52 | |
| 53 | if not isinstance(content, str): |
| 54 | raise TypeError( |
| 55 | f"[{llm_provider}] returned non-text content: {type(content).__name__}" |
| 56 | ) |
| 57 | |
| 58 | # MiniMax M3、DeepSeek R1 这类 reasoning 模型可能会把内部推理包在 |
| 59 | # `<think>...</think>` 中返回。视频脚本和关键词只需要最终可朗读文本， |
| 60 | # 如果不在服务层统一清理，WebUI、字幕和配音都会把思考过程当正文处理。 |
| 61 | content = _THINK_BLOCK_RE.sub("", content) |
| 62 | content = _UNCLOSED_THINK_BLOCK_RE.sub("", content).strip() |
| 63 | if not content: |
| 64 | raise ValueError(f"[{llm_provider}] returned empty text content") |
| 65 | |
| 66 | return content.replace("\n", "") |
| 67 | |
| 68 | |
| 69 | def _sanitize_error_message(error: object) -> str: |
| 70 | """ |
| 71 | 清理返回给 WebUI/API 的错误信息，避免自定义 base_url 中的凭据泄露。 |
| 72 | |
| 73 | 一些 OpenAI-compatible SDK 会把请求 URL 原样拼进异常信息。如果用户为了 |
| 74 | 代理网关配置了 `https://user:pass@example.com/v1`，直接返回 `str(e)` |
| 75 | 就会把密码暴露给页面、API 调用方或后续日志。这里仅处理错误文案，不改变 |
| 76 | 实际请求地址，避免影响正常调用链路。 |
| 77 | """ |
| 78 | message = str(error) |
| 79 | message = _URL_USERINFO_RE.sub(r"\1***:***@", message) |
| 80 | message = _SENSITIVE_QUERY_RE.sub(r"\1***", message) |
| 81 | return message |
| 82 | |
| 83 | |
| 84 | def _extract_chat_completion_text(response, llm_provider: str) -> str: |
| 85 | # OpenAI 兼容接口在异常场景下，可能返回没有 choices、 |
| 86 | # 或者 choices/message/content 为空的响应对象。 |
| 87 | # 这里统一做结构校验，避免出现 `NoneType is not subscriptable` |
| 88 | # 这类底层属性访问错误。 |
| 89 | choices = getattr(response, "choices", None) |
| 90 | if not choices: |
| 91 | raise ValueError(f"[{llm_provider}] returned empty choices") |
| 92 | |
| 93 | first_choice = choices[0] |
| 94 | message = getattr(first_choice, "message", None) |
| 95 | if message is None: |
| 96 | raise ValueError(f"[{llm_provider}] returned empty message") |
| 97 | |
| 98 | content = getattr(message, "content", None) |
| 99 | return _normalize_text_response(content, llm_provider) |
| 100 | |
| 101 | |
| 102 | def _get_response_field(value, key: str): |
| 103 | """兼容 dict 和 SDK 响应对象的字段读取。""" |
| 104 | if isinstance(value, dict): |
| 105 | return value.get(key) |
| 106 | |
| 107 | try: |
| 108 | return value[key] |
| 109 | except (KeyError, TypeError, AttributeError): |
| 110 | return getattr(value, key, None) |
| 111 | |
| 112 | |
| 113 | def _extract_qwen_generation_text(response) -> str: |
| 114 | """ |
| 115 | 从 DashScope Generation 响应中提取文本。 |
| 116 | |
| 117 | Qwen 使用 `messages` 调用时返回的是 chat 结构： |
| 118 | `output.choices[0].message.content`；旧 completion 形态才会返回 |
| 119 | `output.text`。这里两个路径都兼容，避免 `output.text` 为 None 时 |
| 120 | 继续 `.replace()` 触发不可诊断的 AttributeError。 |
| 121 | """ |
| 122 | output = _get_response_field(response, "output") |
| 123 | choices = _get_response_field(output, "choices") if output else None |
| 124 | if choices is not None: |
| 125 | if not choices: |
| 126 | logger.warning("Qwen returned an empty choices list") |
| 127 | raise ValueError("[qwen] returned empty choices") |
| 128 | |
| 129 | first_choice = choices[0] |
| 130 | message = _get_response_field(first_choice, "message") |
| 131 | content = _get_response_field(message, "content") if message else None |
| 132 | if content is not None: |
| 133 | return _normalize_text_response(content, "qwen") |
| 134 | |
| 135 | text = _get_response_field(output, "text") if output else None |
| 136 | return _normalize_text_response(text, "qwen") |
| 137 | |
| 138 | |
| 139 | def _generate_response(prompt: str) -> str: |
| 140 | try: |
| 141 | content = "" |
| 142 | llm_provider = config.app.get("llm_provider", "openai") |
| 143 | logger.info(f"llm provider: {llm_provider}") |
| 144 | if llm_provider == "g4f": |
| 145 | if not config.app.get("enable_g4f", False): |
| 146 | raise ValueError( |
| 147 | "g4f provider is disabled by default because it relies on " |
| 148 | "reverse-engineered third-party endpoints. Set enable_g4f=true " |
| 149 | "in config.toml only if you understand and accept the security, " |
| 150 | "reliability, and legal risks." |
| 151 | ) |
| 152 | |
| 153 | logger.warning( |
| 154 | "g4f provider is enabled. This provider may be unstable and carries " |
| 155 | "supply-chain and terms-of-service risks. Prefer official providers, " |
| 156 | "OpenAI-compatible APIs, LiteLLM, Ollama, or local inference for production." |
| 157 | ) |
| 158 | try: |
| 159 | import g4f |
| 160 | except ImportError as e: |
| 161 | raise ValueError( |
| 162 | "g4f package is not installed by default. Install the optional " |
| 163 | "dependency with `uv sync --extra g4f` only if you understand " |
| 164 | "and accept the provider risks." |
| 165 | ) from e |
| 166 | |
| 167 | model_name = config.app.get("g4f_model_name", "") |
| 168 | if not model_name: |
| 169 | model_name = "gpt-3.5-turbo-16k-0613" |
| 170 | content = g4f.ChatCompletion.create( |
| 171 | model=model_name, |
| 172 | messages=[{"role": "user", "content": prompt}], |
| 173 | ) |
| 174 | else: |
| 175 | api_version = "" # for azure |
| 176 | if llm_provider == "moonshot": |
| 177 | api_key = config.app.get("moonshot_api_key") |
| 178 | model_name = config.app.get("moonshot_model_name") |
| 179 | base_url = "https://api.moonshot.cn/v1" |
| 180 | elif llm_provider == "ollama": |
| 181 | # api_key = config.app.get("openai_api_key") |
| 182 | api_key = "ollama" # any string works but you are required to have one |
| 183 | model_name = config.app.get("ollama_model_name") |
| 184 | base_url = config.app.get("ollama_base_url", "") |
| 185 | if not base_url: |
| 186 | base_url = config.get_default_ollama_base_url() |
| 187 | elif llm_provider == "openai": |
| 188 | api_key = config.app.get("openai_api_key") |
| 189 | model_name = config.app.get("openai_model_name") |
| 190 | base_url = config.app.get("openai_base_url", "") |
| 191 | if not base_url: |
| 192 | base_url = "https://api.openai.com/v1" |
| 193 | elif llm_provider == "aihubmix": |
| 194 | api_key = config.app.get("aihubmix_api_key") |
| 195 | model_name = config.app.get("aihubmix_model_name") |
| 196 | base_url = config.app.get("aihubmix_base_url", "") |
| 197 | # AIHubMix 兼容 OpenAI Chat Completions 协议。这里使用独立 |
| 198 | # provider 保存合作方的默认网关和推荐模型，避免把推广链接、 |
| 199 | # 默认模型等合作配置混进普通 OpenAI provider，影响现有用户。 |
| 200 | if not base_url: |
| 201 | base_url = "https://aihubmix.com/v1" |
| 202 | if not model_name: |
| 203 | model_name = "gpt-5.4-mini" |
| 204 | elif llm_provider == "aimlapi": |
| 205 | api_key = config.app.get("aimlapi_api_key") |
| 206 | model_name = config.app.get("aimlapi_model_name") |
| 207 | base_url = config.app.get("aimlapi_base_url", "") |
| 208 | if not base_url: |
| 209 | base_url = "https://api.aimlapi.com/v1" |
| 210 | if not model_name: |
| 211 | model_name = "openai/gpt-4o-mini" |
| 212 | elif llm_provider == "oneapi": |
| 213 | api_key = config.app.get("oneapi_api_key") |
| 214 | model_name = config.app.get("oneapi_model_name") |
| 215 | base_url = config.app.get("oneapi_base_url", "") |
| 216 | elif llm_provider == "azure": |
| 217 | api_key = config.app.get("azure_api_key") |
| 218 | model_name = config.app.get("azure_model_name") |
| 219 | base_url = config.app.get("azure_base_url", "") |
| 220 | api_version = config.app.get("azure_api_version", "2024-02-15-preview") |
| 221 | elif llm_provider == "gemini": |
| 222 | api_key = config.app.get("gemini_api_key") |
| 223 | model_name = config.app.get("gemini_model_name") |
| 224 | base_url = config.app.get("gemini_base_url", "") |
| 225 | # Gemini 旧模型名已经陆续下线，这里自动兼容历史配置， |
| 226 | # 避免用户沿用旧值时直接收到 404。 |
| 227 | if not model_name: |
| 228 | model_name = _DEFAULT_GEMINI_MODEL |
| 229 | elif model_name in _DEPRECATED_GEMINI_MODELS: |
| 230 | logger.warning( |
| 231 | f"gemini model '{model_name}' is deprecated, fallback to '{_DEFAULT_GEMINI_MODEL}'" |
| 232 | ) |
| 233 | model_name = _DEFAULT_GEMINI_MODEL |
| 234 | elif llm_provider == "grok": |
| 235 | api_key = config.app.get("grok_api_key") |
| 236 | model_name = config.app.get("grok_model_name") |
| 237 | base_url = config.app.get("grok_base_url", "") |
| 238 | if not base_url: |
| 239 | base_url = "https://api.x.ai/v1" |
| 240 | elif llm_provider == "groq": |
| 241 | api_key = config.app.get("groq_api_key") |
| 242 | model_name = config.app.get("groq_model_name") |
| 243 | if not model_name: |
| 244 | model_name = "llama-3.3-70b-versatile" |
| 245 | base_url = config.app.get("groq_base_url", "") |
| 246 | if not base_url: |
| 247 | base_url = "https://api.groq.com/openai/v1" |
| 248 | elif llm_provider == "qwen": |
| 249 | api_key = config.app.get("qwen_api_key") |
| 250 | model_name = config.app.get("qwen_model_name") |
| 251 | base_url = "***" |
| 252 | elif llm_provider == "cloudflare": |
| 253 | api_key = config.app.get("cloudflare_api_key") |
| 254 | model_name = config.app.get("cloudflare_model_name") |
| 255 | account_id = config.app.get("cloudflare_account_id") |
| 256 | base_url = "***" |
| 257 | elif llm_provider == "minimax": |
| 258 | api_key = config.app.get("minimax_api_key") |
| 259 | model_name = config.app.get("minimax_model_name") |
| 260 | base_url = config.app.get("minimax_base_url", "") |
| 261 | if not base_url: |
| 262 | base_url = "https://api.minimax.io/v1" |
| 263 | elif llm_provider == "evolink": |
| 264 | api_key = config.app.get("evolink_api_key") |
| 265 | model_name = config.app.get("evolink_model_name") |
| 266 | base_url = config.app.get("evolink_base_url", "") |
| 267 | if not base_url: |
| 268 | base_url = "https://direct.evolink.ai/v1" |
| 269 | if not model_name: |
| 270 | model_name = "gpt-5.5" |
| 271 | elif llm_provider == "mimo": |
| 272 | api_key = config.app.get("mimo_api_key") |
| 273 | model_name = config.app.get("mimo_model_name") |
| 274 | base_url = config.app.get("mimo_base_url", "") |
| 275 | # Xiaomi MiMo 官方文档说明其兼容 OpenAI Chat Completions 协议。 |
| 276 | # 这里使用独立 provider 保存默认地址和模型名，用户不用把 MiMo |
| 277 | # 当作 OpenAI 自定义 base_url 配置，也便于后续继续接入 MiMo |
| 278 | # 多模态或 TTS 能力时保持边界清晰。 |
| 279 | if not base_url: |
| 280 | base_url = "https://api.xiaomimimo.com/v1" |
| 281 | if not model_name: |
| 282 | model_name = "mimo-v2.5-pro" |
| 283 | elif llm_provider == "volcengine": |
| 284 | api_key = config.app.get("volcengine_api_key") |
| 285 | model_name = config.app.get("volcengine_model_name") |
| 286 | base_url = config.app.get("volcengine_base_url", "") |
| 287 | # 火山引擎方舟提供 OpenAI-compatible Chat Completions 接口。 |
| 288 | # 独立 provider 可以让用户直接选择 VolcEngine，而不用把 Ark |
| 289 | # 的 key/base_url 混到通用 OpenAI 配置里，后续维护也更清晰。 |
| 290 | if not base_url: |
| 291 | base_url = "https://ark.cn-beijing.volces.com/api/v3" |
| 292 | if not model_name: |
| 293 | model_name = "doubao-seed-2-1-turbo-260628" |
| 294 | elif llm_provider == "deepseek": |
| 295 | api_key = config.app.get("deepseek_api_key") |
| 296 | model_name = config.app.get("deepseek_model_name") |
| 297 | base_url = config.app.get("deepseek_base_url") |
| 298 | if not base_url: |
| 299 | base_url = "https://api.deepseek.com" |
| 300 | elif llm_provider == "modelscope": |
| 301 | api_key = config.app.get("modelscope_api_key") |
| 302 | model_name = config.app.get("modelscope_model_name") |
| 303 | base_url = config.app.get("modelscope_base_url") |
| 304 | if not base_url: |
| 305 | base_url = "https://api-inference.modelscope.cn/v1/" |
| 306 | elif llm_provider == "ernie": |
| 307 | api_key = config.app.get("ernie_api_key") |
| 308 | secret_key = config.app.get("ernie_secret_key") |
| 309 | base_url = config.app.get("ernie_base_url") |
| 310 | model_name = "***" |
| 311 | if not secret_key: |
| 312 | raise ValueError( |
| 313 | f"{llm_provider}: secret_key is not set, please set it in the config.toml file." |
| 314 | ) |
| 315 | elif llm_provider == "pollinations": |
| 316 | try: |
| 317 | base_url = config.app.get("pollinations_base_url", "") |
| 318 | if not base_url: |
| 319 | base_url = "https://text.pollinations.ai/openai" |
| 320 | model_name = config.app.get("pollinations_model_name", "openai-fast") |
| 321 | |
| 322 | # Prepare the payload |
| 323 | payload = { |
| 324 | "model": model_name, |
| 325 | "messages": [ |
| 326 | {"role": "user", "content": prompt} |
| 327 | ], |
| 328 | "seed": 101 # Optional but helps with reproducibility |
| 329 | } |
| 330 | |
| 331 | # Optional parameters if configured |
| 332 | if config.app.get("pollinations_private"): |
| 333 | payload["private"] = True |
| 334 | if config.app.get("pollinations_referrer"): |
| 335 | payload["referrer"] = config.app.get("pollinations_referrer") |
| 336 | |
| 337 | headers = { |
| 338 | "Content-Type": "application/json" |
| 339 | } |
| 340 | |
| 341 | # Make the API request |
| 342 | response = requests.post(base_url, headers=headers, json=payload) |
| 343 | response.raise_for_status() |
| 344 | result = response.json() |
| 345 | |
| 346 | if result and "choices" in result and len(result["choices"]) > 0: |
| 347 | content = result["choices"][0]["message"]["content"] |
| 348 | return _normalize_text_response(content, llm_provider) |
| 349 | else: |
| 350 | raise Exception(f"[{llm_provider}] returned an invalid response format") |
| 351 | |
| 352 | except requests.exceptions.RequestException as e: |
| 353 | raise Exception(f"[{llm_provider}] request failed: {str(e)}") |
| 354 | except Exception as e: |
| 355 | raise Exception(f"[{llm_provider}] error: {str(e)}") |
| 356 | |
| 357 | elif llm_provider == "litellm": |
| 358 | model_name = config.app.get("litellm_model_name") |
| 359 | |
| 360 | if llm_provider not in ["pollinations", "ollama", "litellm"]: # Skip validation for providers that don't require API key |
| 361 | if not api_key: |
| 362 | raise ValueError( |
| 363 | f"{llm_provider}: api_key is not set, please set it in the config.toml file." |
| 364 | ) |
| 365 | if not model_name: |
| 366 | raise ValueError( |
| 367 | f"{llm_provider}: model_name is not set, please set it in the config.toml file." |
| 368 | ) |
| 369 | if not base_url and llm_provider not in ["gemini"]: |
| 370 | raise ValueError( |
| 371 | f"{llm_provider}: base_url is not set, please set it in the config.toml file." |
| 372 | ) |
| 373 | |
| 374 | if llm_provider == "qwen": |
| 375 | import dashscope |
| 376 | from dashscope.api_entities.dashscope_response import GenerationResponse |
| 377 | |
| 378 | dashscope.api_key = api_key |
| 379 | response = dashscope.Generation.call( |
| 380 | model=model_name, messages=[{"role": "user", "content": prompt}] |
| 381 | ) |
| 382 | if response: |
| 383 | if isinstance(response, GenerationResponse): |
| 384 | status_code = response.status_code |
| 385 | if status_code != 200: |
| 386 | raise Exception( |
| 387 | f'[{llm_provider}] returned an error response: "{response}"' |
| 388 | ) |
| 389 | |
| 390 | return _extract_qwen_generation_text(response) |
| 391 | else: |
| 392 | raise Exception( |
| 393 | f'[{llm_provider}] returned an invalid response: "{response}"' |
| 394 | ) |
| 395 | else: |
| 396 | raise Exception(f"[{llm_provider}] returned an empty response") |
| 397 | |
| 398 | if llm_provider == "gemini": |
| 399 | import google.generativeai as genai |
| 400 | |
| 401 | if not base_url: |
| 402 | genai.configure(api_key=api_key, transport="rest") |
| 403 | else: |
| 404 | genai.configure(api_key=api_key, transport="rest", client_options={'api_endpoint': base_url}) |
| 405 | |
| 406 | generation_config = { |
| 407 | "temperature": 0.5, |
| 408 | "top_p": 1, |
| 409 | "top_k": 1, |
| 410 | "max_output_tokens": 2048, |
| 411 | } |
| 412 | |
| 413 | safety_settings = [ |
| 414 | { |
| 415 | "category": "HARM_CATEGORY_HARASSMENT", |
| 416 | "threshold": "BLOCK_ONLY_HIGH", |
| 417 | }, |
| 418 | { |
| 419 | "category": "HARM_CATEGORY_HATE_SPEECH", |
| 420 | "threshold": "BLOCK_ONLY_HIGH", |
| 421 | }, |
| 422 | { |
| 423 | "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", |
| 424 | "threshold": "BLOCK_ONLY_HIGH", |
| 425 | }, |
| 426 | { |
| 427 | "category": "HARM_CATEGORY_DANGEROUS_CONTENT", |
| 428 | "threshold": "BLOCK_ONLY_HIGH", |
| 429 | }, |
| 430 | ] |
| 431 | |
| 432 | model = genai.GenerativeModel( |
| 433 | model_name=model_name, |
| 434 | generation_config=generation_config, |
| 435 | safety_settings=safety_settings, |
| 436 | ) |
| 437 | |
| 438 | try: |
| 439 | response = model.generate_content(prompt) |
| 440 | candidates = response.candidates |
| 441 | generated_text = candidates[0].content.parts[0].text |
| 442 | except (AttributeError, IndexError) as e: |
| 443 | logger.warning( |
| 444 | f"gemini returned invalid response content: {str(e)}" |
| 445 | ) |
| 446 | raise ValueError( |
| 447 | f"[{llm_provider}] returned invalid response content" |
| 448 | ) |
| 449 | |
| 450 | return _normalize_text_response(generated_text, llm_provider) |
| 451 | |
| 452 | if llm_provider == "cloudflare": |
| 453 | response = requests.post( |
| 454 | f"https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/{model_name}", |
| 455 | headers={"Authorization": f"Bearer {api_key}"}, |
| 456 | json={ |
| 457 | "messages": [ |
| 458 | { |
| 459 | "role": "system", |
| 460 | "content": "You are a friendly assistant", |
| 461 | }, |
| 462 | {"role": "user", "content": prompt}, |
| 463 | ] |
| 464 | }, |
| 465 | ) |
| 466 | result = response.json() |
| 467 | logger.info(result) |
| 468 | return _normalize_text_response(result["result"]["response"], llm_provider) |
| 469 | |
| 470 | if llm_provider == "ernie": |
| 471 | response = requests.post( |
| 472 | "https://aip.baidubce.com/oauth/2.0/token", |
| 473 | params={ |
| 474 | "grant_type": "client_credentials", |
| 475 | "client_id": api_key, |
| 476 | "client_secret": secret_key, |
| 477 | } |
| 478 | ) |
| 479 | access_token = response.json().get("access_token") |
| 480 | url = f"{base_url}?access_token={access_token}" |
| 481 | |
| 482 | payload = json.dumps( |
| 483 | { |
| 484 | "messages": [{"role": "user", "content": prompt}], |
| 485 | "temperature": 0.5, |
| 486 | "top_p": 0.8, |
| 487 | "penalty_score": 1, |
| 488 | "disable_search": False, |
| 489 | "enable_citation": False, |
| 490 | "response_format": "text", |
| 491 | } |
| 492 | ) |
| 493 | headers = {"Content-Type": "application/json"} |
| 494 | |
| 495 | response = requests.request( |
| 496 | "POST", url, headers=headers, data=payload |
| 497 | ).json() |
| 498 | return _normalize_text_response(response.get("result"), llm_provider) |
| 499 | |
| 500 | if llm_provider == "litellm": |
| 501 | import litellm |
| 502 | |
| 503 | if not model_name: |
| 504 | raise ValueError( |
| 505 | f"{llm_provider}: model_name is not set, please set it in the config.toml file." |
| 506 | ) |
| 507 | |
| 508 | response = litellm.completion( |
| 509 | model=model_name, |
| 510 | messages=[{"role": "user", "content": prompt}], |
| 511 | drop_params=True, |
| 512 | ) |
| 513 | |
| 514 | if not response: |
| 515 | raise ValueError(f"[{llm_provider}] returned empty response") |
| 516 | if not getattr(response, "choices", None): |
| 517 | raise ValueError(f"[{llm_provider}] returned empty response") |
| 518 | |
| 519 | return _extract_chat_completion_text(response, llm_provider) |
| 520 | |
| 521 | if llm_provider == "azure": |
| 522 | # Azure OpenAI SDK 使用 `azure_endpoint` 和 `api_version` 生成专用请求地址， |
| 523 | # 不能继续复用下面普通 OpenAI-compatible 的 `base_url` 初始化逻辑。 |
| 524 | # 这里在 Azure 分支内完成请求并立即返回，避免客户端被后续 fallback |
| 525 | # 覆盖，导致用户配置的 Azure 凭证通过校验但实际请求没有被使用。 |
| 526 | logger.info(f"requesting azure chat completion, model: {model_name}") |
| 527 | client = AzureOpenAI( |
| 528 | api_key=api_key, |
| 529 | api_version=api_version, |
| 530 | azure_endpoint=base_url, |
| 531 | ) |
| 532 | response = client.chat.completions.create( |
| 533 | model=model_name, messages=[{"role": "user", "content": prompt}] |
| 534 | ) |
| 535 | if response: |
| 536 | if isinstance(response, ChatCompletion): |
| 537 | return _extract_chat_completion_text(response, llm_provider) |
| 538 | else: |
| 539 | raise Exception( |
| 540 | f'[{llm_provider}] returned an invalid response: "{response}", please check your network ' |
| 541 | f"connection and try again." |
| 542 | ) |
| 543 | else: |
| 544 | raise Exception( |
| 545 | f"[{llm_provider}] returned an empty response, please check your network connection and try again." |
| 546 | ) |
| 547 | |
| 548 | if llm_provider == "modelscope": |
| 549 | content = '' |
| 550 | client = OpenAI( |
| 551 | api_key=api_key, |
| 552 | base_url=base_url, |
| 553 | ) |
| 554 | response = client.chat.completions.create( |
| 555 | model=model_name, |
| 556 | messages=[{"role": "user", "content": prompt}], |
| 557 | extra_body={"enable_thinking": False}, |
| 558 | stream=True |
| 559 | ) |
| 560 | if response: |
| 561 | for chunk in response: |
| 562 | if not chunk.choices: |
| 563 | continue |
| 564 | delta = chunk.choices[0].delta |
| 565 | if delta and delta.content: |
| 566 | content += delta.content |
| 567 | |
| 568 | if not content.strip(): |
| 569 | raise ValueError("Empty content in stream response") |
| 570 | |
| 571 | return _normalize_text_response(content, llm_provider) |
| 572 | else: |
| 573 | raise Exception(f"[{llm_provider}] returned an empty response") |
| 574 | |
| 575 | else: |
| 576 | client = OpenAI( |
| 577 | api_key=api_key, |
| 578 | base_url=base_url, |
| 579 | ) |
| 580 | |
| 581 | response = client.chat.completions.create( |
| 582 | model=model_name, messages=[{"role": "user", "content": prompt}] |
| 583 | ) |
| 584 | if response: |
| 585 | if isinstance(response, ChatCompletion): |
| 586 | return _extract_chat_completion_text(response, llm_provider) |
| 587 | else: |
| 588 | raise Exception( |
| 589 | f'[{llm_provider}] returned an invalid response: "{response}", please check your network ' |
| 590 | f"connection and try again." |
| 591 | ) |
| 592 | else: |
| 593 | raise Exception( |
| 594 | f"[{llm_provider}] returned an empty response, please check your network connection and try again." |
| 595 | ) |
| 596 | |
| 597 | return _normalize_text_response(content, llm_provider) |
| 598 | except Exception as e: |
| 599 | return f"Error: {_sanitize_error_message(e)}" |
| 600 | |
| 601 | |
| 602 | def _limit_script_text(text: str | None, max_length: int, field_name: str) -> str: |
| 603 | value = (text or "").strip() |
| 604 | if len(value) <= max_length: |
| 605 | return value |
| 606 | |
| 607 | # API 层已经用 Pydantic 做长度校验；这里继续兜底，是为了保护 |
| 608 | # WebUI 或内部服务直接调用 generate_script 时不会把超长提示词发送给模型， |
| 609 | # 避免 token 成本异常和请求失败。 |
| 610 | logger.warning( |
| 611 | f"{field_name} is too long and will be truncated to {max_length} characters." |
| 612 | ) |
| 613 | return value[:max_length] |
| 614 | |
| 615 | |
| 616 | def _normalize_script_paragraph_number(paragraph_number: int | None) -> int: |
| 617 | try: |
| 618 | value = int(paragraph_number or MIN_SCRIPT_PARAGRAPH_NUMBER) |
| 619 | except (TypeError, ValueError): |
| 620 | value = MIN_SCRIPT_PARAGRAPH_NUMBER |
| 621 | |
| 622 | if value < MIN_SCRIPT_PARAGRAPH_NUMBER or value > MAX_SCRIPT_PARAGRAPH_NUMBER: |
| 623 | # WebUI 和 API 都会限制范围；这里兜底处理内部调用，避免异常参数直接扩大 |
| 624 | # LLM 生成成本或生成空结果。 |
| 625 | logger.warning( |
| 626 | "script paragraph_number is out of range and will be clamped: " |
| 627 | f"{value}" |
| 628 | ) |
| 629 | return max(MIN_SCRIPT_PARAGRAPH_NUMBER, min(value, MAX_SCRIPT_PARAGRAPH_NUMBER)) |
| 630 | |
| 631 | return value |
| 632 | |
| 633 | |
| 634 | def build_script_prompt( |
| 635 | video_subject: str, |
| 636 | language: str = "", |
| 637 | paragraph_number: int = 1, |
| 638 | video_script_prompt: str = "", |
| 639 | custom_system_prompt: str = "", |
| 640 | ) -> str: |
| 641 | paragraph_number = _normalize_script_paragraph_number(paragraph_number) |
| 642 | video_script_prompt = _limit_script_text( |
| 643 | video_script_prompt, MAX_SCRIPT_PROMPT_LENGTH, "video_script_prompt" |
| 644 | ) |
| 645 | custom_system_prompt = _limit_script_text( |
| 646 | custom_system_prompt, MAX_SCRIPT_SYSTEM_PROMPT_LENGTH, "custom_system_prompt" |
| 647 | ) |
| 648 | |
| 649 | # 将“脚本生成规则”和“运行时上下文”分开拼接。这样高级用户即使覆盖默认 |
| 650 | # system prompt，也不会漏掉视频主题、语言、段落数这些每次生成都必须带上的参数。 |
| 651 | prompt = custom_system_prompt or DEFAULT_SCRIPT_SYSTEM_PROMPT |
| 652 | prompt += f""" |
| 653 | |
| 654 | # Initialization: |
| 655 | - video subject: {video_subject} |
| 656 | - number of paragraphs: {paragraph_number} |
| 657 | """.rstrip() |
| 658 | if language: |
| 659 | prompt += f"\n- language: {language}" |
| 660 | if video_script_prompt: |
| 661 | prompt += f""" |
| 662 | |
| 663 | # Additional User Requirements: |
| 664 | {video_script_prompt} |
| 665 | """.rstrip() |
| 666 | |
| 667 | return prompt |
| 668 | |
| 669 | |
| 670 | def generate_script( |
| 671 | video_subject: str, |
| 672 | language: str = "", |
| 673 | paragraph_number: int = 1, |
| 674 | video_script_prompt: str = "", |
| 675 | custom_system_prompt: str = "", |
| 676 | ) -> str: |
| 677 | paragraph_number = _normalize_script_paragraph_number(paragraph_number) |
| 678 | video_script_prompt = _limit_script_text( |
| 679 | video_script_prompt, MAX_SCRIPT_PROMPT_LENGTH, "video_script_prompt" |
| 680 | ) |
| 681 | custom_system_prompt = _limit_script_text( |
| 682 | custom_system_prompt, MAX_SCRIPT_SYSTEM_PROMPT_LENGTH, "custom_system_prompt" |
| 683 | ) |
| 684 | prompt = build_script_prompt( |
| 685 | video_subject=video_subject, |
| 686 | language=language, |
| 687 | paragraph_number=paragraph_number, |
| 688 | video_script_prompt=video_script_prompt, |
| 689 | custom_system_prompt=custom_system_prompt, |
| 690 | ) |
| 691 | final_script = "" |
| 692 | logger.info( |
| 693 | "generating video script: " |
| 694 | f"subject={video_subject}, paragraph_number={paragraph_number}, " |
| 695 | f"has_custom_prompt={bool(video_script_prompt.strip())}, " |
| 696 | f"has_custom_system_prompt={bool(custom_system_prompt.strip())}" |
| 697 | ) |
| 698 | |
| 699 | def format_response(response): |
| 700 | # Clean the script |
| 701 | # Remove asterisks, hashes |
| 702 | response = response.replace("*", "") |
| 703 | response = response.replace("#", "") |
| 704 | |
| 705 | # Remove markdown syntax |
| 706 | response = re.sub(r"\[.*\]", "", response) |
| 707 | response = re.sub(r"\(.*\)", "", response) |
| 708 | |
| 709 | # Split the script into paragraphs |
| 710 | paragraphs = response.split("\n\n") |
| 711 | |
| 712 | # Select the specified number of paragraphs |
| 713 | # selected_paragraphs = paragraphs[:paragraph_number] |
| 714 | |
| 715 | # Join the selected paragraphs into a single string |
| 716 | return "\n\n".join(paragraphs) |
| 717 | |
| 718 | for i in range(_max_retries): |
| 719 | try: |
| 720 | response = _generate_response(prompt=prompt) |
| 721 | if response: |
| 722 | final_script = format_response(response) |
| 723 | else: |
| 724 | logging.error("gpt returned an empty response") |
| 725 | |
| 726 | # g4f may return an error message |
| 727 | if final_script and "当日额度已消耗完" in final_script: |
| 728 | raise ValueError(final_script) |
| 729 | |
| 730 | if final_script: |
| 731 | break |
| 732 | except Exception as e: |
| 733 | logger.error(f"failed to generate script: {e}") |
| 734 | |
| 735 | if i < _max_retries: |
| 736 | logger.warning(f"failed to generate video script, trying again... {i + 1}") |
| 737 | if "Error: " in final_script: |
| 738 | logger.error(f"failed to generate video script: {final_script}") |
| 739 | else: |
| 740 | logger.success(f"completed: \n{final_script}") |
| 741 | return final_script.strip() |
| 742 | |
| 743 | |
| 744 | def _strip_code_fence(text: str) -> str: |
| 745 | """Strip a surrounding markdown code fence from an LLM response. |
| 746 | |
| 747 | Non-OpenAI providers (Claude, Gemini, …) frequently wrap JSON output in a |
| 748 | ```json … ``` fence even when asked to return raw JSON. Removing it lets the |
| 749 | first json.loads() succeed instead of falling through to the regex recovery |
| 750 | path (and spuriously logging a warning). Mirrors the DOTALL handling already |
| 751 | used in _parse_social_metadata(). |
| 752 | """ |
| 753 | t = (text or "").strip() |
| 754 | if t.startswith("```"): |
| 755 | t = re.sub(r"^```[a-zA-Z0-9]*\s*", "", t) |
| 756 | t = re.sub(r"\s*```$", "", t) |
| 757 | return t.strip() |
| 758 | |
| 759 | |
| 760 | def generate_terms( |
| 761 | video_subject: str, |
| 762 | video_script: str, |
| 763 | amount: int = 5, |
| 764 | match_script_order: bool = False, |
| 765 | ) -> List[str]: |
| 766 | if match_script_order: |
| 767 | goal = ( |
| 768 | f"Generate {amount} chronological stock-video search terms that follow " |
| 769 | "the order of topics in the video script." |
| 770 | ) |
| 771 | ordering_rule = ( |
| 772 | "6. keep the terms in the same order as the script narration; " |
| 773 | "earlier terms must describe earlier visual moments." |
| 774 | ) |
| 775 | # 有序关键词模式下，示例数量要和 amount 保持一致，避免模型被固定 |
| 776 | # 的 4 个示例误导，导致长文案只返回少量关键词，影响素材覆盖度。 |
| 777 | example_terms = [ |
| 778 | "opening visual topic", |
| 779 | *[ |
| 780 | f"script visual topic {index}" |
| 781 | for index in range(2, max(amount, 1)) |
| 782 | ], |
| 783 | "final visual topic", |
| 784 | ] |
| 785 | output_example = json.dumps(example_terms[:amount], ensure_ascii=False) |
| 786 | else: |
| 787 | goal = ( |
| 788 | f"Generate {amount} search terms for stock videos, depending on the " |
| 789 | "subject of a video." |
| 790 | ) |
| 791 | ordering_rule = "" |
| 792 | output_example = ( |
| 793 | '["search term 1", "search term 2", "search term 3",' |
| 794 | '"search term 4", "search term 5"]' |
| 795 | ) |
| 796 | |
| 797 | prompt = f""" |
| 798 | # Role: Video Search Terms Generator |
| 799 | |
| 800 | ## Goals: |
| 801 | {goal} |
| 802 | |
| 803 | ## Constrains: |
| 804 | 1. the search terms are to be returned as a json-array of strings. |
| 805 | 2. each search term should consist of 1-3 words, always add the main subject of the video. |
| 806 | 3. you must only return the json-array of strings. you must not return anything else. you must not return the script. |
| 807 | 4. the search terms must be related to the subject of the video. |
| 808 | 5. reply with english search terms only. |
| 809 | {ordering_rule} |
| 810 | |
| 811 | ## Output Example: |
| 812 | {output_example} |
| 813 | |
| 814 | ## Context: |
| 815 | ### Video Subject |
| 816 | {video_subject} |
| 817 | |
| 818 | ### Video Script |
| 819 | {video_script} |
| 820 | |
| 821 | Please note that you must use English for generating video search terms; Chinese is not accepted. |
| 822 | """.strip() |
| 823 | |
| 824 | logger.info( |
| 825 | f"subject: {video_subject}, match_script_order: {match_script_order}" |
| 826 | ) |
| 827 | |
| 828 | search_terms = [] |
| 829 | response = "" |
| 830 | for i in range(_max_retries): |
| 831 | try: |
| 832 | response = _generate_response(prompt) |
| 833 | if "Error: " in response: |
| 834 | logger.error(f"failed to generate video script: {response}") |
| 835 | return response |
| 836 | search_terms = json.loads(_strip_code_fence(response)) |
| 837 | if not isinstance(search_terms, list) or not all( |
| 838 | isinstance(term, str) for term in search_terms |
| 839 | ): |
| 840 | logger.error("response is not a list of strings.") |
| 841 | continue |
| 842 | |
| 843 | except Exception as e: |
| 844 | logger.warning(f"failed to generate video terms: {str(e)}") |
| 845 | if response: |
| 846 | match = re.search(r"\[.*]", response, re.DOTALL) |
| 847 | if match: |
| 848 | try: |
| 849 | search_terms = json.loads(match.group()) |
| 850 | except Exception as e: |
| 851 | # 这里保留重试流程，但必须记录 LLM 返回的非标准 JSON， |
| 852 | # 否则后续排查搜索词为空时无法定位 |
| 853 | # 是模型格式问题还是解析逻辑问题。 |
| 854 | logger.warning(f"failed to generate video terms: {str(e)}") |
| 855 | |
| 856 | if search_terms and len(search_terms) > 0: |
| 857 | break |
| 858 | if i < _max_retries: |
| 859 | logger.warning(f"failed to generate video terms, trying again... {i + 1}") |
| 860 | |
| 861 | logger.success(f"completed: \n{search_terms}") |
| 862 | return search_terms |
| 863 | |
| 864 | |
| 865 | # ============================================================================= |
| 866 | # Social publishing metadata |
| 867 | # |
| 868 | # 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。 |
| 869 | # 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。 |
| 870 | # ============================================================================= |
| 871 | |
| 872 | # 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 |
| 873 | # 过长内容后调用方还需要二次裁剪。 |
| 874 | SOCIAL_PLATFORMS = { |
| 875 | "tiktok": {"title_max": 100, "caption_max": 2200, "hashtag_count": 5}, |
| 876 | "youtube_shorts": {"title_max": 100, "caption_max": 5000, "hashtag_count": 3}, |
| 877 | "instagram_reels": {"title_max": 125, "caption_max": 2200, "hashtag_count": 8}, |
| 878 | "facebook_reels": {"title_max": 125, "caption_max": 2200, "hashtag_count": 5}, |
| 879 | } |
| 880 | DEFAULT_SOCIAL_PLATFORM = "tiktok" |
| 881 | DEFAULT_SOCIAL_LANGUAGE = "auto" |
| 882 | MAX_SOCIAL_SUBJECT_LENGTH = 500 |
| 883 | MAX_SOCIAL_SCRIPT_LENGTH = 8000 |
| 884 | MAX_SOCIAL_LANGUAGE_LENGTH = 64 |
| 885 | |
| 886 | SOCIAL_PLATFORM_LABELS = { |
| 887 | "tiktok": "TikTok", |
| 888 | "youtube_shorts": "YouTube Shorts", |
| 889 | "instagram_reels": "Instagram Reels", |
| 890 | "facebook_reels": "Facebook Reels", |
| 891 | } |
| 892 | |
| 893 | # 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 |
| 894 | # 对中文、英文、越南语等不同场景都能返回可用结构。 |
| 895 | DEFAULT_SOCIAL_HASHTAGS = [ |
| 896 | "#shorts", |
| 897 | "#viral", |
| 898 | "#trending", |
| 899 | "#fyp", |
| 900 | "#video", |
| 901 | "#reels", |
| 902 | "#creator", |
| 903 | "#content", |
| 904 | ] |
| 905 | |
| 906 | |
| 907 | def _resolve_social_platform(platform: str | None) -> str: |
| 908 | value = (platform or "").strip().lower() |
| 909 | return value if value in SOCIAL_PLATFORMS else DEFAULT_SOCIAL_PLATFORM |
| 910 | |
| 911 | |
| 912 | def _normalize_social_language(language: str | None) -> str: |
| 913 | value = (language or DEFAULT_SOCIAL_LANGUAGE).strip() |
| 914 | if len(value) > MAX_SOCIAL_LANGUAGE_LENGTH: |
| 915 | logger.warning( |
| 916 | "social metadata language is too long and will be truncated to " |
| 917 | f"{MAX_SOCIAL_LANGUAGE_LENGTH} characters." |
| 918 | ) |
| 919 | value = value[:MAX_SOCIAL_LANGUAGE_LENGTH] |
| 920 | return value or DEFAULT_SOCIAL_LANGUAGE |
| 921 | |
| 922 | |
| 923 | def _limit_social_text(text: str | None, max_length: int, field_name: str) -> str: |
| 924 | value = (text or "").strip() |
| 925 | if len(value) <= max_length: |
| 926 | return value |
| 927 | |
| 928 | # 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 |
| 929 | # 直接调用时不会把超长内容发送给模型，避免 token 成本异常。 |
| 930 | logger.warning( |
| 931 | f"{field_name} is too long and will be truncated to {max_length} characters." |
| 932 | ) |
| 933 | return value[:max_length] |
| 934 | |
| 935 | |
| 936 | def _social_language_instruction(language: str | None) -> str: |
| 937 | language = _normalize_social_language(language) |
| 938 | if language.lower() == DEFAULT_SOCIAL_LANGUAGE: |
| 939 | return ( |
| 940 | "Use the same language as the video subject and script. If the subject " |
| 941 | "and script use different languages, prefer the script language." |
| 942 | ) |
| 943 | |
| 944 | return f'Write "title" and "caption" in this language: {language}.' |
| 945 | |
| 946 | |
| 947 | def _clamp_text(text, max_length: int) -> str: |
| 948 | value = ("" if text is None else str(text)).strip() |
| 949 | if max_length and len(value) > max_length: |
| 950 | return value[:max_length].rstrip() |
| 951 | return value |
| 952 | |
| 953 | |
| 954 | def _normalize_hashtags(raw, count: int) -> List[str]: |
| 955 | """ |
| 956 | 将 LLM 返回的 hashtag 统一整理成 `#tag` 格式。 |
| 957 | |
| 958 | 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。 |
| 959 | 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、 |
| 960 | 重复标签或不符合常见格式的 hashtag。 |
| 961 | """ |
| 962 | if isinstance(raw, str): |
| 963 | candidates = re.split(r"[\s,]+", raw) |
| 964 | elif isinstance(raw, (list, tuple)): |
| 965 | # 数组里的每一项视为一个完整标签，因此 "du lich" 会变成 |
| 966 | # "#dulich"，而不是拆成两个标签。 |
| 967 | candidates = [str(entry) for entry in raw] |
| 968 | else: |
| 969 | candidates = [] |
| 970 | |
| 971 | seen = set() |
| 972 | result: List[str] = [] |
| 973 | for item in candidates: |
| 974 | tag = re.sub(r"[^\w]", "", item, flags=re.UNICODE) |
| 975 | if not tag: |
| 976 | continue |
| 977 | key = tag.lower() |
| 978 | if key in seen: |
| 979 | continue |
| 980 | seen.add(key) |
| 981 | result.append(f"#{tag}") |
| 982 | if count and len(result) >= count: |
| 983 | break |
| 984 | return result |
| 985 | |
| 986 | |
| 987 | def build_social_metadata_prompt( |
| 988 | video_subject: str, |
| 989 | video_script: str = "", |
| 990 | language: str = DEFAULT_SOCIAL_LANGUAGE, |
| 991 | platform: str = DEFAULT_SOCIAL_PLATFORM, |
| 992 | ) -> str: |
| 993 | video_subject = _limit_social_text( |
| 994 | video_subject, MAX_SOCIAL_SUBJECT_LENGTH, "video_subject" |
| 995 | ) |
| 996 | video_script = _limit_social_text( |
| 997 | video_script, MAX_SOCIAL_SCRIPT_LENGTH, "video_script" |
| 998 | ) |
| 999 | platform = _resolve_social_platform(platform) |
| 1000 | spec = SOCIAL_PLATFORMS[platform] |
| 1001 | label = SOCIAL_PLATFORM_LABELS.get(platform, platform) |
| 1002 | language_instruction = _social_language_instruction(language) |
| 1003 | |
| 1004 | prompt = f""" |
| 1005 | # Role: Short-Video Social Media Copywriter |
| 1006 | |
| 1007 | ## Goal |
| 1008 | Write engaging publishing metadata for a short video that will be posted on {label}. |
| 1009 | |
| 1010 | ## Constraints |
| 1011 | 1. Respond ONLY with a single valid minified JSON object. No markdown, no code fences, no commentary. |
| 1012 | 2. The JSON must contain exactly these keys: "title", "caption", "hashtags". |
| 1013 | 3. "title": a catchy hook, at most {spec['title_max']} characters. |
| 1014 | 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. |
| 1015 | 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}. |
| 1016 | 6. {language_instruction} |
| 1017 | |
| 1018 | ## Output Example |
| 1019 | {{"title":"...","caption":"...","hashtags":["#example","#video"]}} |
| 1020 | |
| 1021 | ## Context |
| 1022 | ### Video Subject |
| 1023 | {video_subject} |
| 1024 | |
| 1025 | ### Video Script |
| 1026 | {video_script} |
| 1027 | """.strip() |
| 1028 | return prompt |
| 1029 | |
| 1030 | |
| 1031 | def _parse_social_metadata(response: str, platform: str) -> dict: |
| 1032 | spec = SOCIAL_PLATFORMS[_resolve_social_platform(platform)] |
| 1033 | |
| 1034 | data = None |
| 1035 | try: |
| 1036 | data = json.loads(_strip_code_fence(response)) |
| 1037 | except Exception: |
| 1038 | # 部分模型会在 JSON 外层包一段说明文字或 markdown fence。 |
| 1039 | # API 调用方只需要稳定结构，所以这里尝试提取第一个 JSON object。 |
| 1040 | match = re.search(r"\{.*\}", response or "", re.DOTALL) |
| 1041 | if match: |
| 1042 | data = json.loads(match.group()) |
| 1043 | |
| 1044 | if not isinstance(data, dict): |
| 1045 | raise ValueError("social metadata response is not a JSON object") |
| 1046 | |
| 1047 | title = _clamp_text(data.get("title", ""), spec["title_max"]) |
| 1048 | caption = _clamp_text(data.get("caption", ""), spec["caption_max"]) |
| 1049 | hashtags = _normalize_hashtags(data.get("hashtags", []), spec["hashtag_count"]) |
| 1050 | |
| 1051 | if not title and not caption: |
| 1052 | raise ValueError("social metadata response is missing both title and caption") |
| 1053 | |
| 1054 | return {"title": title, "caption": caption, "hashtags": hashtags} |
| 1055 | |
| 1056 | |
| 1057 | def _fallback_social_metadata( |
| 1058 | video_subject: str, video_script: str, platform: str |
| 1059 | ) -> dict: |
| 1060 | spec = SOCIAL_PLATFORMS[_resolve_social_platform(platform)] |
| 1061 | subject = (video_subject or "").strip() |
| 1062 | script = (video_script or "").strip() |
| 1063 | |
| 1064 | title = subject |
| 1065 | if not title and script: |
| 1066 | # 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。 |
| 1067 | title = re.split(r"(?<=[.!?。！？])\s+", script)[0] |
| 1068 | |
| 1069 | return { |
| 1070 | "title": _clamp_text(title, spec["title_max"]), |
| 1071 | "caption": _clamp_text(script or subject, spec["caption_max"]), |
| 1072 | "hashtags": _normalize_hashtags( |
| 1073 | DEFAULT_SOCIAL_HASHTAGS, spec["hashtag_count"] |
| 1074 | ), |
| 1075 | } |
| 1076 | |
| 1077 | |
| 1078 | def generate_social_metadata( |
| 1079 | video_subject: str, |
| 1080 | video_script: str = "", |
| 1081 | language: str = DEFAULT_SOCIAL_LANGUAGE, |
| 1082 | platform: str = DEFAULT_SOCIAL_PLATFORM, |
| 1083 | ) -> dict: |
| 1084 | """ |
| 1085 | 生成短视频发布文案元数据。 |
| 1086 | |
| 1087 | 返回结构固定为 `{"title": str, "caption": str, "hashtags": List[str]}`。 |
| 1088 | 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 |
| 1089 | 调用方始终拿到可展示、可发布前编辑的数据结构。 |
| 1090 | """ |
| 1091 | platform = _resolve_social_platform(platform) |
| 1092 | language = _normalize_social_language(language) |
| 1093 | video_subject = _limit_social_text( |
| 1094 | video_subject, MAX_SOCIAL_SUBJECT_LENGTH, "video_subject" |
| 1095 | ) |
| 1096 | video_script = _limit_social_text( |
| 1097 | video_script, MAX_SOCIAL_SCRIPT_LENGTH, "video_script" |
| 1098 | ) |
| 1099 | prompt = build_social_metadata_prompt( |
| 1100 | video_subject=video_subject, |
| 1101 | video_script=video_script, |
| 1102 | language=language, |
| 1103 | platform=platform, |
| 1104 | ) |
| 1105 | logger.info( |
| 1106 | f"generating social metadata: platform={platform}, language={language}" |
| 1107 | ) |
| 1108 | |
| 1109 | response = "" |
| 1110 | for i in range(_max_retries): |
| 1111 | try: |
| 1112 | response = _generate_response(prompt) |
| 1113 | if isinstance(response, str) and "Error: " in response: |
| 1114 | logger.error(f"failed to generate social metadata: {response}") |
| 1115 | break |
| 1116 | metadata = _parse_social_metadata(response, platform) |
| 1117 | logger.success(f"completed: \n{metadata}") |
| 1118 | return metadata |
| 1119 | except Exception as e: |
| 1120 | logger.warning(f"failed to parse social metadata: {str(e)}") |
| 1121 | |
| 1122 | if i < _max_retries - 1: |
| 1123 | logger.warning( |
| 1124 | f"failed to generate social metadata, trying again... {i + 1}" |
| 1125 | ) |
| 1126 | |
| 1127 | logger.warning("falling back to heuristic social metadata") |
| 1128 | return _fallback_social_metadata(video_subject, video_script, platform) |
| 1129 | |
| 1130 | |
| 1131 | if __name__ == "__main__": |
| 1132 | video_subject = "生命的意义是什么" |
| 1133 | script = generate_script( |
| 1134 | video_subject=video_subject, language="zh-CN", paragraph_number=1 |
| 1135 | ) |
| 1136 | print("######################") |
| 1137 | print(script) |
| 1138 | search_terms = generate_terms( |
| 1139 | video_subject=video_subject, video_script=script, amount=5 |
| 1140 | ) |
| 1141 | print("######################") |
| 1142 | print(search_terms) |
| 1143 | |
| 1144 |