wonder
f3abc5fdc6
Deploy Slides / build-and-deploy (push) Successful in 1m39s
- eino-deep-dive.md: split slide 2 into code-only + explanation+SVG slides - async-deep-dive.md: split slide 2 into code-only + progress+SVG slides - harness.md: split slide 2 (4 Items + SVG) into two slides (2 Items each) - request-journey.md: split into middleware/handler + pipeline/finalization slides The frankfurt theme header (~40px) overlaps ~7px into the content zone, leaving ~467px effective height. Dense slides with code+Items+images exceeded this budget and were hidden behind the bars.
2.2 KiB
2.2 KiB
Eino Graph 深入 — 构建与编排
有向图构建与本地状态管理
g := compose.NewGraph[PipelineInput, PipelineOutput](
compose.WithGenLocalState(func(ctx context.Context) *PipelineState {
return &PipelineState{} // 每次调用获得全新状态
}),
)
g.AddLambdaNode(nodePromptOptimizer, &compose.Lambda{
StatePreHandler: preparePromptInput, // 从 state 读取输入
StatePostHandler: storePromptOutput, // 结果写回 state
})
每次 `graph.Invoke()` 调用生成独立的 `PipelineState`，避免并发调用间的状态污染。State 在节点间通过 Pre/Post Handler 流转。
PreHandler 从全局 State 提取当前节点所需字段；PostHandler 将节点输出写回 State。节点本身无状态，所有数据经 State 中转。
Eino Graph 深入 — 分支路由与重试
QualitySupervisor 驱动的条件分支
g.AddBranch(nodeQualitySupervisor, compose.NewGraphBranch(
func(ctx context.Context, _ PipelineInput) (string, error) {
var next string
_ = compose.ProcessState[*PipelineState](ctx,
func(_ context.Context, state *PipelineState) error {
next = state.NextNode // 质检结果决定路由
return nil
})
return next, nil
},
map[string]bool{
nodePromptOptimizer: true,
nodeFormatAdapter: true,
},
))
分支路由 — 逻辑与全貌
质检通过 → FormatAdapter 直接输出；不通过且 retry < 3 → PromptOptimizer 重新优化；不通过且 retry >= 3 → FormatAdapter 降级输出。 下图展示了完整的分支路由与重试机制，包括状态流转和降级路径。