e9d7f8832a
- Backend: FastAPI + SQLAlchemy + Celery (Python 3.11+) - Frontend: Vue 3 + TypeScript + Pinia + Tailwind - Admin Frontend: separate Vue 3 app for management - Docker Compose: 9 services orchestration - Specs: design prototypes, memory system PRD, product roadmap Cleanup performed: - Removed temporary debug scripts from backend root - Removed deprecated admin_app.py (embedded UI) - Removed duplicate docs from admin-frontend - Updated .gitignore for Vite cache and egg-info
678 lines
23 KiB
Markdown
678 lines
23 KiB
Markdown
# RFC: 供应商平台化架构设计
|
||
|
||
## 背景
|
||
|
||
### 当前问题
|
||
1. **硬编码适配器**: `gemini`, `flux`, `minimax` 写死在代码中
|
||
2. **新供应商需改代码**: 接入 nanobanana 等新供应商需要修改 `provider_router.py`
|
||
3. **无法动态切换**: 供应商故障时需要重启服务
|
||
4. **缺乏监控**: 不知道哪个供应商更快、更便宜、更稳定
|
||
|
||
### 目标
|
||
- **零代码接入**: 通过后台配置即可接入新供应商
|
||
- **动态切换**: 运行时切换供应商，无需重启
|
||
- **智能路由**: 基于成本、延迟、成功率自动选择最优供应商
|
||
- **可观测性**: 供应商健康状态、成本、性能一目了然
|
||
|
||
---
|
||
|
||
## 架构设计
|
||
|
||
### 1. 整体架构
|
||
|
||
```
|
||
┌─────────────────────────────────────────────────────────────┐
|
||
│ Admin Dashboard │
|
||
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────┐ │
|
||
│ │ 供应商管理 │ │ 健康监控 │ │ 成本分析 │ │
|
||
│ └─────────────┘ └─────────────┘ └─────────────────────┘ │
|
||
└─────────────────────────────────────────────────────────────┘
|
||
│
|
||
▼
|
||
┌─────────────────────────────────────────────────────────────┐
|
||
│ Provider Router │
|
||
│ ┌─────────────────────────────────────────────────────┐ │
|
||
│ │ 路由策略: Priority → Weight → Health → Cost │ │
|
||
│ └─────────────────────────────────────────────────────┘ │
|
||
│ │ │
|
||
│ ┌───────────┬───────────┬───────────┬───────────────────┐ │
|
||
│ │ Adapter │ Adapter │ Adapter │ Adapter │ │
|
||
│ │ Registry │ Factory │ Health │ Metrics │ │
|
||
│ └───────────┴───────────┴───────────┴───────────────────┘ │
|
||
└─────────────────────────────────────────────────────────────┘
|
||
│
|
||
┌─────────────────────┼─────────────────────┐
|
||
▼ ▼ ▼
|
||
┌───────────────┐ ┌───────────────┐ ┌───────────────┐
|
||
│ Text Adapters │ │ Image Adapters│ │ TTS Adapters │
|
||
├───────────────┤ ├───────────────┤ ├───────────────┤
|
||
│ • Gemini │ │ • Flux │ │ • Minimax │
|
||
│ • OpenAI │ │ • Nanobanana │ │ • ElevenLabs │
|
||
│ • Claude │ │ • DALL-E │ │ • Azure TTS │
|
||
│ • Qwen │ │ • Midjourney │ │ • Google TTS │
|
||
└───────────────┘ └───────────────┘ └───────────────┘
|
||
```
|
||
|
||
### 2. 核心组件
|
||
|
||
#### 2.1 Adapter 接口定义
|
||
|
||
```python
|
||
# 统一适配器接口
|
||
from abc import ABC, abstractmethod
|
||
from typing import TypeVar, Generic
|
||
from pydantic import BaseModel
|
||
|
||
T = TypeVar("T")
|
||
|
||
class AdapterConfig(BaseModel):
|
||
"""适配器配置基类"""
|
||
api_key: str
|
||
api_base: str | None = None
|
||
model: str | None = None
|
||
timeout_ms: int = 60000
|
||
max_retries: int = 3
|
||
|
||
class BaseAdapter(ABC, Generic[T]):
|
||
"""适配器基类"""
|
||
|
||
# 适配器元信息
|
||
adapter_type: str # text / image / tts
|
||
adapter_name: str # gemini / flux / minimax
|
||
|
||
def __init__(self, config: AdapterConfig):
|
||
self.config = config
|
||
|
||
@abstractmethod
|
||
async def execute(self, **kwargs) -> T:
|
||
"""执行适配器逻辑"""
|
||
pass
|
||
|
||
@abstractmethod
|
||
async def health_check(self) -> bool:
|
||
"""健康检查"""
|
||
pass
|
||
|
||
@property
|
||
@abstractmethod
|
||
def estimated_cost(self) -> float:
|
||
"""预估单次调用成本 (USD)"""
|
||
pass
|
||
```
|
||
|
||
#### 2.2 适配器注册表
|
||
|
||
```python
|
||
# 适配器注册表 - 支持动态注册
|
||
class AdapterRegistry:
|
||
"""适配器注册表"""
|
||
|
||
_adapters: dict[str, type[BaseAdapter]] = {}
|
||
|
||
@classmethod
|
||
def register(cls, adapter_type: str, adapter_name: str):
|
||
"""装饰器: 注册适配器"""
|
||
def decorator(adapter_class: type[BaseAdapter]):
|
||
key = f"{adapter_type}:{adapter_name}"
|
||
cls._adapters[key] = adapter_class
|
||
return adapter_class
|
||
return decorator
|
||
|
||
@classmethod
|
||
def get(cls, adapter_type: str, adapter_name: str) -> type[BaseAdapter] | None:
|
||
key = f"{adapter_type}:{adapter_name}"
|
||
return cls._adapters.get(key)
|
||
|
||
@classmethod
|
||
def list_adapters(cls, adapter_type: str | None = None) -> list[str]:
|
||
"""列出所有已注册的适配器"""
|
||
if adapter_type:
|
||
return [k for k in cls._adapters if k.startswith(f"{adapter_type}:")]
|
||
return list(cls._adapters.keys())
|
||
```
|
||
|
||
#### 2.3 适配器实现示例
|
||
|
||
```python
|
||
# 图像适配器示例: Nanobanana
|
||
@AdapterRegistry.register("image", "nanobanana")
|
||
class NanobananapAdapter(BaseAdapter[str]):
|
||
adapter_type = "image"
|
||
adapter_name = "nanobanana"
|
||
|
||
async def execute(self, prompt: str, **kwargs) -> str:
|
||
"""生成图片，返回 URL"""
|
||
async with httpx.AsyncClient(timeout=self.config.timeout_ms / 1000) as client:
|
||
response = await client.post(
|
||
f"{self.config.api_base}/generate",
|
||
json={"prompt": prompt, "model": self.config.model},
|
||
headers={"Authorization": f"Bearer {self.config.api_key}"},
|
||
)
|
||
response.raise_for_status()
|
||
return response.json()["image_url"]
|
||
|
||
async def health_check(self) -> bool:
|
||
# 简单的健康检查
|
||
try:
|
||
async with httpx.AsyncClient(timeout=5) as client:
|
||
response = await client.get(f"{self.config.api_base}/health")
|
||
return response.status_code == 200
|
||
except Exception:
|
||
return False
|
||
|
||
@property
|
||
def estimated_cost(self) -> float:
|
||
return 0.02 # $0.02 per image
|
||
```
|
||
|
||
#### 2.4 智能路由器
|
||
|
||
```python
|
||
class ProviderRouter:
|
||
"""智能供应商路由器"""
|
||
|
||
def __init__(self, db: AsyncSession):
|
||
self.db = db
|
||
self._health_cache: dict[str, tuple[bool, float]] = {} # adapter_key -> (healthy, last_check)
|
||
|
||
async def route(
|
||
self,
|
||
provider_type: str,
|
||
strategy: str = "priority", # priority / cost / latency / round_robin
|
||
**kwargs
|
||
):
|
||
"""路由到最优供应商"""
|
||
providers = await self._get_enabled_providers(provider_type)
|
||
|
||
if not providers:
|
||
raise ValueError(f"No {provider_type} providers configured")
|
||
|
||
# 按策略排序
|
||
sorted_providers = self._sort_by_strategy(providers, strategy)
|
||
|
||
errors = []
|
||
for provider in sorted_providers:
|
||
# 检查健康状态
|
||
if not await self._is_healthy(provider):
|
||
continue
|
||
|
||
try:
|
||
adapter = self._create_adapter(provider)
|
||
result = await adapter.execute(**kwargs)
|
||
|
||
# 记录成功指标
|
||
await self._record_metrics(provider, success=True)
|
||
return result
|
||
|
||
except Exception as e:
|
||
errors.append(f"{provider.name}: {e}")
|
||
await self._record_metrics(provider, success=False, error=str(e))
|
||
continue
|
||
|
||
raise ValueError(f"All providers failed: {' | '.join(errors)}")
|
||
|
||
def _sort_by_strategy(self, providers: list[Provider], strategy: str) -> list[Provider]:
|
||
if strategy == "priority":
|
||
return sorted(providers, key=lambda p: (-p.priority, -p.weight))
|
||
elif strategy == "cost":
|
||
return sorted(providers, key=lambda p: self._get_estimated_cost(p))
|
||
elif strategy == "latency":
|
||
return sorted(providers, key=lambda p: self._get_avg_latency(p))
|
||
else:
|
||
return providers
|
||
```
|
||
|
||
### 3. 数据模型扩展
|
||
|
||
```sql
|
||
-- 供应商表 (已有，需扩展)
|
||
ALTER TABLE providers ADD COLUMN api_key_ref VARCHAR(100); -- 密钥引用 (从 secrets 表获取)
|
||
ALTER TABLE providers ADD COLUMN request_schema JSONB; -- 请求参数 schema
|
||
ALTER TABLE providers ADD COLUMN response_parser VARCHAR(200); -- 响应解析规则
|
||
|
||
-- 供应商指标表 (新增)
|
||
CREATE TABLE provider_metrics (
|
||
id SERIAL PRIMARY KEY,
|
||
provider_id VARCHAR(36) REFERENCES providers(id),
|
||
timestamp TIMESTAMP WITH TIME ZONE DEFAULT NOW(),
|
||
success BOOLEAN,
|
||
latency_ms INTEGER,
|
||
cost_usd DECIMAL(10, 6),
|
||
error_message TEXT,
|
||
request_id VARCHAR(100)
|
||
);
|
||
|
||
-- 供应商健康状态表 (新增)
|
||
CREATE TABLE provider_health (
|
||
provider_id VARCHAR(36) PRIMARY KEY REFERENCES providers(id),
|
||
is_healthy BOOLEAN DEFAULT TRUE,
|
||
last_check TIMESTAMP WITH TIME ZONE,
|
||
consecutive_failures INTEGER DEFAULT 0,
|
||
last_error TEXT
|
||
);
|
||
|
||
-- 密钥管理表 (新增)
|
||
CREATE TABLE provider_secrets (
|
||
id VARCHAR(36) PRIMARY KEY,
|
||
name VARCHAR(100) UNIQUE NOT NULL,
|
||
encrypted_value TEXT NOT NULL, -- 加密存储
|
||
created_at TIMESTAMP WITH TIME ZONE DEFAULT NOW(),
|
||
updated_at TIMESTAMP WITH TIME ZONE DEFAULT NOW()
|
||
);
|
||
```
|
||
|
||
### 4. Admin Dashboard 功能
|
||
|
||
#### 4.1 供应商管理
|
||
- 供应商列表 (启用/禁用/删除)
|
||
- 新增供应商 (选择适配器类型 + 配置参数)
|
||
- 编辑供应商 (修改优先级/权重/超时等)
|
||
- 测试连接 (验证 API Key 有效性)
|
||
|
||
#### 4.2 健康监控
|
||
- 实时健康状态 (绿/黄/红)
|
||
- 成功率趋势图
|
||
- 延迟分布图
|
||
- 故障告警配置
|
||
|
||
#### 4.3 成本分析
|
||
- 按供应商统计调用量
|
||
- 按供应商统计成本
|
||
- 成本趋势图
|
||
- 预算告警
|
||
|
||
#### 4.4 A/B 测试
|
||
- 创建实验 (供应商 A vs B)
|
||
- 流量分配 (50/50 或自定义)
|
||
- 效果对比 (成功率/延迟/成本)
|
||
|
||
---
|
||
|
||
## 实现路径
|
||
|
||
### 阶段 1: 适配器抽象 (基础) - ✅ 已完成
|
||
|
||
| 任务 | 状态 | 文件 |
|
||
|------|------|------|
|
||
| 定义 `BaseAdapter` 接口 | ✅ | `services/adapters/base.py` |
|
||
| 实现 `AdapterRegistry` 注册表 | ✅ | `services/adapters/registry.py` |
|
||
| 重构 GeminiAdapter | ✅ | `services/adapters/text/gemini.py` |
|
||
| 重构 FluxAdapter | ✅ | `services/adapters/image/flux.py` |
|
||
| 重构 MinimaxAdapter | ✅ | `services/adapters/tts/minimax.py` |
|
||
| 重构 `ProviderRouter` 使用新接口 | ✅ | `services/provider_router.py` |
|
||
|
||
### 阶段 2: 新供应商接入 (扩展) - 待开始
|
||
1. 实现 Nanobanana 适配器
|
||
2. 实现 OpenAI/Claude 文本适配器
|
||
3. 实现 ElevenLabs TTS 适配器
|
||
4. 验证零代码接入流程
|
||
|
||
### 阶段 3: 监控与分析 (可观测) - 待开始
|
||
1. 实现指标收集
|
||
2. 实现健康检查
|
||
3. 实现成本追踪
|
||
4. Admin Dashboard 开发
|
||
|
||
### 阶段 4: 智能路由 (优化) - 待开始
|
||
1. 实现多种路由策略
|
||
2. 实现自动故障转移
|
||
3. 实现 A/B 测试框架
|
||
|
||
---
|
||
|
||
## 并行执行与容错设计
|
||
|
||
### 问题
|
||
|
||
当前串行流程存在两个问题：
|
||
1. **等待时间长**: 故事(3-5s) → 封面(5-10s) → 音频(3-5s) = 总计 11-20s
|
||
2. **单点失败**: 某一步502/超时导致整个流程失败
|
||
|
||
### 方案 1: 并行执行
|
||
|
||
```python
|
||
async def generate_story_full(keywords: list[str]) -> StoryResult:
|
||
# Step 1: 故事生成（必须先完成，后续依赖它）
|
||
story = await generate_story_content(keywords)
|
||
|
||
# Step 2: 图片和音频并行执行
|
||
image_task = asyncio.create_task(generate_image(story.summary))
|
||
audio_task = asyncio.create_task(text_to_speech(story.content))
|
||
|
||
# 等待两者完成，互不阻塞
|
||
image_result, audio_result = await asyncio.gather(
|
||
image_task, audio_task,
|
||
return_exceptions=True # 一个<E4B880><E4B8AA><EFBFBD>败不影响另一个
|
||
)
|
||
|
||
return StoryResult(
|
||
story=story,
|
||
image_url=image_result if not isinstance(image_result, Exception) else None,
|
||
audio_url=audio_result if not isinstance(audio_result, Exception) else None,
|
||
errors={
|
||
"image": str(image_result) if isinstance(image_result, Exception) else None,
|
||
"audio": str(audio_result) if isinstance(audio_result, Exception) else None,
|
||
}
|
||
)
|
||
```
|
||
|
||
**时间对比:**
|
||
```
|
||
串行: 3s + 8s + 4s = 15s
|
||
并行: 3s + max(8s, 4s) = 11s (节省 27%)
|
||
```
|
||
|
||
### 方案 2: 部分成功处理
|
||
|
||
**核心原则: 部分成功 > 全部失败**
|
||
|
||
```python
|
||
@dataclass
|
||
class StoryResult:
|
||
story: Story # 核心，必须成功
|
||
image_url: str | None = None # 增强，可降级
|
||
audio_url: str | None = None # 增强，可降级
|
||
errors: dict[str, str] = field(default_factory=dict)
|
||
|
||
@property
|
||
def is_complete(self) -> bool:
|
||
return self.image_url is not None and self.audio_url is not None
|
||
|
||
@property
|
||
def failed_components(self) -> list[str]:
|
||
return [k for k, v in self.errors.items() if v is not None]
|
||
```
|
||
|
||
**降级策略:**
|
||
|
||
| 组件 | 失败时降级方案 | 用户体验 |
|
||
|------|---------------|---------|
|
||
| 故事 | 无降级，整体失败 | 显示错误，提示重试 |
|
||
| 封面 | 使用默认封面图 | 显示占位图 + "重新生成"按钮 |
|
||
| 音频 | 不生成音频 | 隐藏播放按钮 + "生成语音"按钮 |
|
||
|
||
### 方案 3: 流式返回 (SSE)
|
||
|
||
**为什么用 SSE:**
|
||
- 用户无需等待全部完成
|
||
- 每完成一步立即展示
|
||
- 比 WebSocket 简单，HTTP 兼容性好
|
||
|
||
**后端实现:**
|
||
|
||
```python
|
||
from fastapi import APIRouter
|
||
from sse_starlette.sse import EventSourceResponse
|
||
|
||
router = APIRouter()
|
||
|
||
@router.post("/api/generate/stream")
|
||
async def generate_story_stream(
|
||
request: GenerateRequest,
|
||
current_user: User = Depends(get_current_user),
|
||
):
|
||
async def event_generator():
|
||
# 1. 立即返回任务ID
|
||
story_id = str(uuid.uuid4())
|
||
yield {"event": "started", "data": json.dumps({"story_id": story_id})}
|
||
|
||
# 2. 生成故事
|
||
try:
|
||
story = await generate_story_content(request.keywords)
|
||
yield {"event": "story_ready", "data": json.dumps({
|
||
"title": story.title,
|
||
"content": story.content,
|
||
})}
|
||
except Exception as e:
|
||
yield {"event": "story_failed", "data": json.dumps({"error": str(e)})}
|
||
return
|
||
|
||
# 3. 并行生成图片和音频
|
||
async def gen_image():
|
||
try:
|
||
url = await generate_image(story.summary)
|
||
yield {"event": "image_ready", "data": json.dumps({"image_url": url})}
|
||
except Exception as e:
|
||
yield {"event": "image_failed", "data": json.dumps({"error": str(e)})}
|
||
|
||
async def gen_audio():
|
||
try:
|
||
url = await text_to_speech(story.content)
|
||
yield {"event": "audio_ready", "data": json.dumps({"audio_url": url})}
|
||
except Exception as e:
|
||
yield {"event": "audio_failed", "data": json.dumps({"error": str(e)})}
|
||
|
||
# 并行执行，逐个yield结果
|
||
tasks = [gen_image(), gen_audio()]
|
||
for coro in asyncio.as_completed([t.__anext__() for t in tasks]):
|
||
result = await coro
|
||
yield result
|
||
|
||
yield {"event": "complete", "data": json.dumps({"story_id": story_id})}
|
||
|
||
return EventSourceResponse(event_generator())
|
||
```
|
||
|
||
**前端实现:**
|
||
|
||
```typescript
|
||
const eventSource = new EventSource('/api/generate/stream', {
|
||
method: 'POST',
|
||
body: JSON.stringify({ keywords }),
|
||
});
|
||
|
||
eventSource.addEventListener('started', (e) => {
|
||
const { story_id } = JSON.parse(e.data);
|
||
showLoading('正在创作故事...');
|
||
});
|
||
|
||
eventSource.addEventListener('story_ready', (e) => {
|
||
const { title, content } = JSON.parse(e.data);
|
||
renderStory(title, content);
|
||
showLoading('正在生成封面和语音...');
|
||
});
|
||
|
||
eventSource.addEventListener('image_ready', (e) => {
|
||
const { image_url } = JSON.parse(e.data);
|
||
renderCover(image_url);
|
||
});
|
||
|
||
eventSource.addEventListener('image_failed', (e) => {
|
||
showRetryButton('image');
|
||
});
|
||
|
||
eventSource.addEventListener('audio_ready', (e) => {
|
||
const { audio_url } = JSON.parse(e.data);
|
||
enablePlayButton(audio_url);
|
||
});
|
||
|
||
eventSource.addEventListener('complete', () => {
|
||
eventSource.close();
|
||
hideLoading();
|
||
});
|
||
```
|
||
|
||
**用户体验时间线:**
|
||
```
|
||
0s → 显示"正在创作..."
|
||
3s → 故事文本渲染，显示"正在生成封面和语音..."
|
||
3-7s → 音频就绪，播放按钮可用
|
||
3-11s → 封面就绪，图片显示
|
||
11s → 完成
|
||
```
|
||
|
||
### 方案 4: 断点续传 (可选)
|
||
|
||
适用于网络不稳定场景，支持刷新页面后继续：
|
||
|
||
```python
|
||
class StoryWorkflowState(Base):
|
||
__tablename__ = "story_workflow_states"
|
||
|
||
story_id: Mapped[str] = mapped_column(String(36), primary_key=True)
|
||
status: Mapped[str] = mapped_column(String(20)) # pending/story_done/image_done/audio_done/complete
|
||
story_content: Mapped[str | None] = mapped_column(Text)
|
||
image_url: Mapped[str | None] = mapped_column(String(500))
|
||
audio_url: Mapped[str | None] = mapped_column(String(500))
|
||
last_error: Mapped[str | None] = mapped_column(Text)
|
||
created_at: Mapped[datetime] = mapped_column(DateTime, default=datetime.utcnow)
|
||
updated_at: Mapped[datetime] = mapped_column(DateTime, onupdate=datetime.utcnow)
|
||
|
||
async def resume_workflow(story_id: str) -> StoryResult:
|
||
state = await get_workflow_state(story_id)
|
||
|
||
if state.status == "story_done":
|
||
# 从图片+音频生成继续
|
||
return await generate_image_and_audio(state)
|
||
elif state.status == "image_done":
|
||
# 只需要生成音频
|
||
return await generate_audio_only(state)
|
||
elif state.status == "audio_done":
|
||
# 只需要生成图片
|
||
return await generate_image_only(state)
|
||
else:
|
||
return StoryResult.from_state(state)
|
||
```
|
||
|
||
### 推荐实现顺序
|
||
|
||
| 优先级 | 方案 | 收益 | 复杂度 | 状态 |
|
||
|--------|------|------|--------|------|
|
||
| P0 | 并行执行 | 节省 27% 时间 | 低 | ✅ 已完成 |
|
||
| P0 | 部分成功 | 提升容错性 | 低 | ✅ 已完成 |
|
||
| P1 | SSE 流式返回 | 体验大幅提升 | 中 | 待开始 |
|
||
| P2 | 断点续传 | 极端场景保障 | 高 | 待开始 |
|
||
|
||
**P0 实现详情:**
|
||
- 新增 API: `POST /api/generate/full`
|
||
- 文件: `api/stories.py:113-189`
|
||
- 响应模型: `FullStoryResponse` (含 `errors` 字段标识失败组件)
|
||
|
||
---
|
||
|
||
## 待决策清单
|
||
|
||
> **使用说明**: 在每个决策的 `[ ]` 中填入你的选择（如 `[x]` 或 `[B]`），确认后删除未选中的选项。
|
||
|
||
---
|
||
|
||
### 决策 1: 适配器配置存储
|
||
|
||
**问题**: 适配器的配置信息（API地址、模型名、超时等）存在哪里？
|
||
|
||
| 选项 | 方案 | 优点 | 缺点 |
|
||
|------|------|------|------|
|
||
| [ ] A | 全部存数据库 | 完全动态，运行时可改 | 需要管理界面，初始化复杂 |
|
||
| [ ] B | 代码定义 + DB配置 | 平衡，核心逻辑在代码，参数可调 | 新适配器仍需改代码 |
|
||
| [ ] C | 配置文件 (YAML/JSON) | 简单，版本控制友好 | 改配置需重启 |
|
||
|
||
**推荐**: B（代码定义适配器类，DB存储启用状态/优先级/API Key引用）
|
||
|
||
---
|
||
|
||
### 决策 2: 密钥管理
|
||
|
||
**问题**: API Key 等敏感信息如何存储？
|
||
|
||
| 选项 | 方案 | 优点 | 缺点 |
|
||
|------|------|------|------|
|
||
| [ ] A | 环境变量 | 简单，当前方式 | 多供应商时env膨胀，改key需重启 |
|
||
| [ ] B | 数据库加密存储 | 动态管理，支持多key | 需要加密方案，安全风险 |
|
||
| [ ] C | 外部密钥服务 (Vault/AWS Secrets) | 企业级安全 | 复杂，增加依赖 |
|
||
|
||
**推荐**: A（当前阶段），后期可迁移到B
|
||
|
||
---
|
||
|
||
### 决策 3: 图像供应商优先级
|
||
|
||
**问题**: 接入多个图像供应商后，默认使用哪个？
|
||
|
||
| 选项 | 供应商 | 特点 | 预估成本 |
|
||
|------|--------|------|----------|
|
||
| [ ] 1 | Nanobanana | 新兴，据说效果好 | 待调研 |
|
||
| [ ] 2 | Flux (当前) | 稳定，已接入 | ~$0.03/张 |
|
||
| [ ] 3 | DALL-E 3 | OpenAI出品，质量高 | ~$0.04/张 |
|
||
| [ ] 4 | Midjourney | 艺术风格强 | API受限 |
|
||
|
||
**推荐**: 先调研Nanobanana，效果好则替换Flux
|
||
|
||
---
|
||
|
||
### 决策 4: 文本供应商优先级
|
||
|
||
**问题**: 故事生成使用哪个LLM？
|
||
|
||
| 选项 | 供应商 | 特点 | 预估成本 |
|
||
|------|--------|------|----------|
|
||
| [ ] 1 | Gemini (当前) | 免费额度大，中文好 | 免费/低成本 |
|
||
| [ ] 2 | OpenAI GPT-4o | 质量稳定 | ~$0.01/1K tokens |
|
||
| [ ] 3 | Claude | 创意写作强 | ~$0.015/1K tokens |
|
||
| [ ] 4 | Qwen (通义千问) | 国内，中文优化 | 待调研 |
|
||
|
||
**推荐**: Gemini为主，OpenAI备用
|
||
|
||
---
|
||
|
||
### 决策 5: TTS供应商优先级
|
||
|
||
**问题**: 语音合成使用哪个服务？
|
||
|
||
| 选项 | 供应商 | 特点 | 预估成本 |
|
||
|------|--------|------|----------|
|
||
| [ ] 1 | Minimax (当前) | 中文效果好，已接入 | ~$0.01/1K字符 |
|
||
| [ ] 2 | ElevenLabs | 英文最佳，多语言 | ~$0.03/1K字符 |
|
||
| [ ] 3 | Azure TTS | 稳定，多语言 | ~$0.016/1K字符 |
|
||
| [ ] 4 | Google TTS | 便宜 | ~$0.004/1K字符 |
|
||
|
||
**推荐**: Minimax为主（中文场景）
|
||
|
||
---
|
||
|
||
### 决策 6: Admin Dashboard 技术栈
|
||
|
||
**问题**: 供应商管理后台用什么技术？
|
||
|
||
| 选项 | 方案 | 优点 | 缺点 |
|
||
|------|------|------|------|
|
||
| [ ] A | 复用 Vue 前端 | 技术栈统一，复用组件 | 需要自己写UI |
|
||
| [ ] B | React Admin | 成熟的Admin框架 | 引入新技术栈 |
|
||
| [ ] C | 现成方案 (AdminJS/Retool) | 开发快 | 定制性差，可能收费 |
|
||
|
||
**推荐**: A（在现有Vue项目中加 `/admin` 路由）
|
||
|
||
---
|
||
|
||
### 决策 7: Phase 2 功能优先级
|
||
|
||
**问题**: 体验增强阶段先做哪个功能？
|
||
|
||
| 选项 | 功能 | 用户价值 | 开发复杂度 |
|
||
|------|------|----------|------------|
|
||
| [ ] 1 | 故事编辑 | 高（用户可修改AI内容） | 中 |
|
||
| [ ] 2 | 角色定制 | 高（孩子成为主角） | 低 |
|
||
| [ ] 3 | 故事分享 | 高（增长引擎） | 中 |
|
||
| [ ] 4 | 故事续写 | 中（延长使用时长） | 中 |
|
||
|
||
**推荐**: 2 → 1 → 3 → 4（角色定制最快出效果）
|
||
|
||
---
|
||
|
||
### 决策 8: 并行与容错实现顺序
|
||
|
||
**问题**: 并行执行、部分成功、SSE、断点续传，先做哪些？
|
||
|
||
| 选项 | 方案 | 说明 |
|
||
|------|------|------|
|
||
| [ ] A | P0先做 | 先实现并行+部分成功，快速见效 |
|
||
| [ ] B | P0+P1一起 | 并行+部分成功+SSE，体验完整 |
|
||
| [ ] C | 只做SSE | 跳过简单方案，直接上流式 |
|
||
|
||
**推荐**: A（先P0，验证后再做SSE）
|
||
|
||
---
|
||
|
||
## 确认后删除此区块
|
||
|
||
确认所有决策后，可以删除未选中的选项，保留最终方案作为实现依据。
|