285 lines
12 KiB
TypeScript
285 lines
12 KiB
TypeScript
|
import { buildApiUrl, buildAuthHeaders } from "./serverConnection";
|
||
|
|
|
||
|
|
const TEXT_MODEL = "qwen-max";
|
||
|
|
const VISION_MODEL = "qwen3.6-plus";
|
||
|
|
|
||
|
|
export interface AdVideoUserConfig {
|
||
|
|
platform: string;
|
||
|
|
aspectRatio: string;
|
||
|
|
durationSeconds: number;
|
||
|
|
style: string;
|
||
|
|
language: string;
|
||
|
|
market: string;
|
||
|
|
needVoiceover: boolean;
|
||
|
|
needSubtitle: boolean;
|
||
|
|
conversionFocus: "conversion" | "brand";
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface ProductSummary {
|
||
|
|
product_name: string;
|
||
|
|
category: string;
|
||
|
|
appearance: string;
|
||
|
|
materials: string[];
|
||
|
|
colors: string[];
|
||
|
|
core_features: string[];
|
||
|
|
target_users: string[];
|
||
|
|
usage_scenarios: string[];
|
||
|
|
selling_points: string[];
|
||
|
|
risk_notes: string[];
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface SellingPoint {
|
||
|
|
point: string;
|
||
|
|
evidence: string;
|
||
|
|
ad_expression: string;
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface SellingPointResult {
|
||
|
|
primary_selling_points: SellingPoint[];
|
||
|
|
secondary_selling_points: SellingPoint[];
|
||
|
|
unsupported_claims: string[];
|
||
|
|
compliance_warnings: string[];
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface CreativeOption {
|
||
|
|
creative_id: string;
|
||
|
|
creative_type: string;
|
||
|
|
hook: string;
|
||
|
|
target_user: string;
|
||
|
|
main_message: string;
|
||
|
|
emotional_tone: string;
|
||
|
|
recommended_platform: string;
|
||
|
|
reason: string;
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface StoryboardScene {
|
||
|
|
scene_id: number;
|
||
|
|
duration: string;
|
||
|
|
scene_goal: string;
|
||
|
|
visual_description: string;
|
||
|
|
product_focus: string;
|
||
|
|
camera_movement: string;
|
||
|
|
background: string;
|
||
|
|
lighting: string;
|
||
|
|
subtitle: string;
|
||
|
|
voiceover: string;
|
||
|
|
transition: string;
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface Storyboard {
|
||
|
|
video_title: string;
|
||
|
|
duration: string;
|
||
|
|
aspect_ratio: string;
|
||
|
|
target_platform: string;
|
||
|
|
language: string;
|
||
|
|
scenes: StoryboardScene[];
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface VideoPrompt {
|
||
|
|
scene_id: number;
|
||
|
|
positive_prompt: string;
|
||
|
|
negative_prompt: string;
|
||
|
|
reference_requirements: string;
|
||
|
|
consistency_rules: string;
|
||
|
|
text_overlay: string;
|
||
|
|
}
|
||
|
|
|
||
|
|
export interface ComplianceCheck {
|
||
|
|
risk_level: "low" | "medium" | "high";
|
||
|
|
issues: Array<{ field: string; problem: string; suggestion: string }>;
|
||
|
|
allow_video_generation: boolean;
|
||
|
|
}
|
||
|
|
|
||
|
|
function extractJson(text: string): unknown {
|
||
|
|
const fenced = text.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||
|
|
const raw = fenced ? fenced[1].trim() : text.trim();
|
||
|
|
const start = raw.search(/[[{]/);
|
||
|
|
const slice = start >= 0 ? raw.slice(start) : raw;
|
||
|
|
try {
|
||
|
|
return JSON.parse(slice);
|
||
|
|
} catch {
|
||
|
|
throw new Error("AI 返回内容不是有效的 JSON");
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
interface ChatMessage {
|
||
|
|
role: "system" | "user";
|
||
|
|
content: string;
|
||
|
|
}
|
||
|
|
|
||
|
|
async function chat(
|
||
|
|
systemPrompt: string,
|
||
|
|
userContent: string,
|
||
|
|
options?: { model?: string; signal?: AbortSignal },
|
||
|
|
): Promise<string> {
|
||
|
|
const messages: ChatMessage[] = [
|
||
|
|
{ role: "system", content: systemPrompt },
|
||
|
|
{ role: "user", content: userContent },
|
||
|
|
];
|
||
|
|
const timeoutSignal = AbortSignal.timeout(60000);
|
||
|
|
const combinedSignal = options?.signal
|
||
|
|
? AbortSignal.any([options.signal, timeoutSignal])
|
||
|
|
: timeoutSignal;
|
||
|
|
const res = await fetch(buildApiUrl("ai/chat"), {
|
||
|
|
method: "POST",
|
||
|
|
headers: buildAuthHeaders(),
|
||
|
|
body: JSON.stringify({
|
||
|
|
model: options?.model ?? TEXT_MODEL,
|
||
|
|
messages,
|
||
|
|
stream: false,
|
||
|
|
temperature: 0.4,
|
||
|
|
}),
|
||
|
|
signal: combinedSignal,
|
||
|
|
});
|
||
|
|
if (!res.ok) throw new Error(`AI 调用失败 (${res.status})`);
|
||
|
|
const payload = await res.json();
|
||
|
|
const content: string =
|
||
|
|
payload?.content ?? payload?.choices?.[0]?.message?.content ?? payload?.text ?? "";
|
||
|
|
if (!content) throw new Error("模型未返回有效内容");
|
||
|
|
return content;
|
||
|
|
}
|
||
|
|
|
||
|
|
async function visionChat(
|
||
|
|
systemPrompt: string,
|
||
|
|
text: string,
|
||
|
|
imageUrls: string[],
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<string> {
|
||
|
|
const content = [
|
||
|
|
...imageUrls.map((url) => ({ type: "image_url", image_url: { url } })),
|
||
|
|
{ type: "text", text },
|
||
|
|
];
|
||
|
|
const timeoutSignal = AbortSignal.timeout(60000);
|
||
|
|
const combinedSignal = signal
|
||
|
|
? AbortSignal.any([signal, timeoutSignal])
|
||
|
|
: timeoutSignal;
|
||
|
|
const res = await fetch(buildApiUrl("ai/chat"), {
|
||
|
|
method: "POST",
|
||
|
|
headers: buildAuthHeaders(),
|
||
|
|
body: JSON.stringify({
|
||
|
|
model: VISION_MODEL,
|
||
|
|
messages: [
|
||
|
|
{ role: "system", content: systemPrompt },
|
||
|
|
{ role: "user", content },
|
||
|
|
],
|
||
|
|
stream: false,
|
||
|
|
temperature: 0.3,
|
||
|
|
}),
|
||
|
|
signal: combinedSignal,
|
||
|
|
});
|
||
|
|
if (!res.ok) throw new Error(`图片理解调用失败 (${res.status})`);
|
||
|
|
const payload = await res.json();
|
||
|
|
const out: string =
|
||
|
|
payload?.content ?? payload?.choices?.[0]?.message?.content ?? payload?.text ?? "";
|
||
|
|
if (!out) throw new Error("图片理解未返回有效内容");
|
||
|
|
return out;
|
||
|
|
}
|
||
|
|
|
||
|
|
const IMAGE_UNDERSTANDING_PROMPT = `你是电商产品图片分析专家。请分析用户提供的产品图片，识别产品主体、外观、颜色、材质、形状、尺寸感、品牌标识、关键部件、可视化卖点和适合展示的镜头角度。请用简洁的中文段落描述，不要编造图片中看不到的信息。`;
|
||
|
|
|
||
|
|
export async function analyzeProductImages(
|
||
|
|
imageUrls: string[],
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<string> {
|
||
|
|
if (imageUrls.length === 0) return "";
|
||
|
|
return visionChat(IMAGE_UNDERSTANDING_PROMPT, "请分析这些产品图片的视觉特征。", imageUrls, signal);
|
||
|
|
}
|
||
|
|
|
||
|
|
const PRODUCT_SUMMARY_PROMPT = `你是商品信息理解专家。根据产品图片理解结果和说明书文本，输出结构化的商品信息。严格按以下 JSON 格式返回，不要任何额外解释：
|
||
|
|
{"product_name":"","category":"","appearance":"","materials":[],"colors":[],"core_features":[],"target_users":[],"usage_scenarios":[],"selling_points":[],"risk_notes":[]}
|
||
|
|
要求：只描述资料中真实存在的信息，不要编造说明书或图片中不存在的功能。risk_notes 列出可能涉及夸大、医疗功效、绝对化用语等风险点。`;
|
||
|
|
|
||
|
|
export async function buildProductSummary(
|
||
|
|
imageDescription: string,
|
||
|
|
manualText: string,
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<ProductSummary> {
|
||
|
|
const userContent = `【产品图片理解结果】\n${imageDescription || "（无图片）"}\n\n【产品说明书/详情文本】\n${manualText || "（无文本）"}`;
|
||
|
|
const text = await chat(PRODUCT_SUMMARY_PROMPT, userContent, { signal });
|
||
|
|
return extractJson(text) as ProductSummary;
|
||
|
|
}
|
||
|
|
|
||
|
|
const SELLING_POINT_PROMPT = `你是电商卖点提炼专家。将商品信息拆分为不同层级卖点。严格按以下 JSON 格式返回，不要任何额外解释：
|
||
|
|
{"primary_selling_points":[{"point":"","evidence":"","ad_expression":""}],"secondary_selling_points":[{"point":"","evidence":"","ad_expression":""}],"unsupported_claims":[],"compliance_warnings":[]}
|
||
|
|
要求：每个卖点必须有来源依据(evidence)，依据来自输入的商品信息。不得凭空增加功能。无依据的卖点放入 unsupported_claims。涉及夸大、医疗、绝对化用语的放入 compliance_warnings。`;
|
||
|
|
|
||
|
|
export async function extractSellingPoints(
|
||
|
|
summary: ProductSummary,
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<SellingPointResult> {
|
||
|
|
const text = await chat(SELLING_POINT_PROMPT, `【商品结构化信息】\n${JSON.stringify(summary, null, 2)}`, { signal });
|
||
|
|
return extractJson(text) as SellingPointResult;
|
||
|
|
}
|
||
|
|
|
||
|
|
function configBlock(config: AdVideoUserConfig): string {
|
||
|
|
return `【用户配置】\n目标平台：${config.platform}\n视频比例：${config.aspectRatio}\n时长：${config.durationSeconds}秒\n广告风格：${config.style}\n语言：${config.language}\n目标市场：${config.market}\n旁白：${config.needVoiceover ? "需要" : "不需要"}\n字幕：${config.needSubtitle ? "需要" : "不需要"}\n侧重：${config.conversionFocus === "conversion" ? "强转化" : "品牌展示"}`;
|
||
|
|
}
|
||
|
|
|
||
|
|
const CREATIVE_PROMPT = `你是电商广告创意专家。根据商品卖点和用户配置，生成至少 3 个差异化的广告创意方向。严格按以下 JSON 格式返回，不要任何额外解释：
|
||
|
|
{"creative_options":[{"creative_id":"A","creative_type":"","hook":"","target_user":"","main_message":"","emotional_tone":"","recommended_platform":"","reason":""}]}
|
||
|
|
要求：每个方向围绕真实卖点，有清晰广告逻辑，方向之间有明显差异。`;
|
||
|
|
|
||
|
|
export async function generateCreativeOptions(
|
||
|
|
selling: SellingPointResult,
|
||
|
|
config: AdVideoUserConfig,
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<CreativeOption[]> {
|
||
|
|
const userContent = `【卖点】\n${JSON.stringify(selling.primary_selling_points, null, 2)}\n\n${configBlock(config)}`;
|
||
|
|
const text = await chat(CREATIVE_PROMPT, userContent, { signal });
|
||
|
|
const parsed = extractJson(text) as { creative_options?: CreativeOption[] };
|
||
|
|
return Array.isArray(parsed.creative_options) ? parsed.creative_options : [];
|
||
|
|
}
|
||
|
|
|
||
|
|
const STORYBOARD_PROMPT = `你是电商短视频分镜师。根据选定的广告创意方向、商品信息和用户配置，输出结构化视频分镜。严格按以下 JSON 格式返回，不要任何额外解释：
|
||
|
|
{"video_title":"","duration":"","aspect_ratio":"","target_platform":"","language":"","scenes":[{"scene_id":1,"duration":"3s","scene_goal":"","visual_description":"","product_focus":"","camera_movement":"","background":"","lighting":"","subtitle":"","voiceover":"","transition":""}]}
|
||
|
|
要求：开头3秒有吸引点，中段展示核心卖点，结尾有行动号召。各镜头时长之和等于配置总时长。不要出现说明书中不存在的功能，不要设计视频模型难以稳定生成的复杂动作。`;
|
||
|
|
|
||
|
|
export async function generateStoryboard(
|
||
|
|
creative: CreativeOption,
|
||
|
|
summary: ProductSummary,
|
||
|
|
config: AdVideoUserConfig,
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<Storyboard> {
|
||
|
|
const userContent = `【选定创意方向】\n${JSON.stringify(creative, null, 2)}\n\n【商品信息】\n${JSON.stringify(summary, null, 2)}\n\n${configBlock(config)}`;
|
||
|
|
const text = await chat(STORYBOARD_PROMPT, userContent, { signal });
|
||
|
|
return extractJson(text) as Storyboard;
|
||
|
|
}
|
||
|
|
|
||
|
|
const VIDEO_PROMPT_PROMPT = `你是 AI 视频模型提示词工程师。为每个分镜生成视频模型提示词。严格按以下 JSON 格式返回一个数组，不要任何额外解释：
|
||
|
|
[{"scene_id":1,"positive_prompt":"","negative_prompt":"","reference_requirements":"","consistency_rules":"","text_overlay":""}]
|
||
|
|
正向提示词需包含：产品主体、外观、颜色、材质、使用场景、镜头构图、镜头运动、光线风格、背景环境、广告质感、画面节奏。
|
||
|
|
负面提示词需包含：不改变产品外观/颜色、不添加不存在的部件、不生成错误Logo、不生成模糊文字、不生成虚假功能演示、不生成畸形手部、不生成夸张功效、不生成医学暗示。
|
||
|
|
字幕和文字建议后期叠加(text_overlay)，不要让视频模型直接生成文字。`;
|
||
|
|
|
||
|
|
export async function generateVideoPrompts(
|
||
|
|
storyboard: Storyboard,
|
||
|
|
summary: ProductSummary,
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<VideoPrompt[]> {
|
||
|
|
const userContent = `【分镜脚本】\n${JSON.stringify(storyboard.scenes, null, 2)}\n\n【产品外观特征(一致性参考)】\n外观：${summary.appearance}\n颜色：${summary.colors.join("、")}\n材质：${summary.materials.join("、")}`;
|
||
|
|
const text = await chat(VIDEO_PROMPT_PROMPT, userContent, { signal });
|
||
|
|
const parsed = extractJson(text);
|
||
|
|
return Array.isArray(parsed) ? (parsed as VideoPrompt[]) : [];
|
||
|
|
}
|
||
|
|
|
||
|
|
const COMPLIANCE_PROMPT = `你是电商广告合规质检专家。检查文案和卖点是否存在虚假宣传、绝对化用语(如"最""第一""100%")、医疗功效暗示、高风险品类违规表达。严格按以下 JSON 格式返回，不要任何额外解释：
|
||
|
|
{"risk_level":"low","issues":[{"field":"","problem":"","suggestion":""}],"allow_video_generation":true}
|
||
|
|
risk_level 取值 low/medium/high。存在高风险违规时 allow_video_generation 设为 false。`;
|
||
|
|
|
||
|
|
export async function checkCompliance(
|
||
|
|
summary: ProductSummary,
|
||
|
|
selling: SellingPointResult,
|
||
|
|
storyboard: Storyboard,
|
||
|
|
signal?: AbortSignal,
|
||
|
|
): Promise<ComplianceCheck> {
|
||
|
|
const userContent = `【卖点】\n${JSON.stringify(selling, null, 2)}\n\n【分镜文案/旁白/字幕】\n${JSON.stringify(storyboard.scenes.map((s) => ({ subtitle: s.subtitle, voiceover: s.voiceover })), null, 2)}\n\n【风险点】\n${summary.risk_notes.join("、")}`;
|
||
|
|
const text = await chat(COMPLIANCE_PROMPT, userContent, { signal });
|
||
|
|
return extractJson(text) as ComplianceCheck;
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|