wonder
d479504f12
Deploy Slides / build-and-deploy (push) Successful in 1m32s
- Restructure from 12 slides to 21 slides with proper section navigation - Split into pages/ for modular maintenance (15 page files) - Use frankfurt theme: cover layout, Item component, section frontmatter - Remove all two-cols layouts to avoid rendering issues - Add rich content from tmp/ docs: Eino pipeline, prompt agent, async tasks, data model, deployment - Use Item component for structured content blocks - Add mermaid ER diagram for data model - Add Go code block for pipeline type definitions
30 lines
1.1 KiB
Markdown
30 lines
1.1 KiB
Markdown
# 提示词优化 Agent
|
|
|
|
Eino Chain 驱动的智能提示词生成
|
|
|
|
```mermaid {scale: 0.8}
|
|
graph LR
|
|
A[用户描述] --> B[风格合并]
|
|
B --> C[三段式组装]
|
|
C --> D[LLM 优化]
|
|
D --> E[最终提示词]
|
|
F[工程风格] --> B
|
|
G[任务覆盖] --> B
|
|
H[重试注入] --> D
|
|
```
|
|
|
|
<Item title="① 风格合并">
|
|
`finalStyle = merge(projectStyle, taskStyle)` — 工程风格为基底，任务风格逐键覆盖。自定义标签（`custom:` 前缀）自动提取。
|
|
</Item>
|
|
|
|
<Item title="② 三段式提示词组装">
|
|
将合并后的风格信息组装为三段式规范提示词：
|
|
- **主体描述**：用户输入的角色/场景描述
|
|
- **风格约束**：美术风格、色调、线条等风格键
|
|
- **技术规格**：尺寸、格式、透明背景等引擎适配参数
|
|
</Item>
|
|
|
|
<Item title="③ LLM 优化 + 回退机制">
|
|
调用 Eino Chain 将三段式提示词发给 DeepSeek 等 LLM 进行润色优化。**无 API Key 时自动回退到模板生成**，保证功能可用。优化失败时注入重试上下文，指导模型修正。
|
|
</Item>
|