d3cec2a5b9
✨ feat(agent): 新增智能排程 Agent 全链路 + ReAct 精排引擎 🏗️ 智能排程 Graph 编排（阶段 1 基础链路） - 新增 scheduleplan 包：state / tool / prompt / nodes / runner / graph 六件套 - 实现 plan → preview → materialize → apply → reflect → finalize 完整图编排 - 通过函数注入解耦 agent 层与 service 层，避免循环依赖 - 路由层新增 schedule_plan 动作，复用现有 SSE + 持久化链路 🧠 ReAct 精排引擎（阶段 1.5 语义化微调） - 粗排后构建"混合日程"（既有课程 + 建议任务），统一为 HybridScheduleEntry - LLM 开启深度思考，通过 Swap / Move / TimeAvailable / GetAvailableSlots 四个 Tool 在内存中优化任务时间 - reasoning_content 实时流式推送前端，用户可见 AI 思考过程 - 精排结果仅预览不落库，向后兼容（未注入依赖时走原有 materialize 路径） 📝 文档 - 新增 ReAct 精排引擎决策记录 ⚠️ 已知问题：深度思考模式耗时较长，超时策略待优化
244 lines
8.3 KiB
Go
244 lines
8.3 KiB
Go
package scheduleplan
|
||
|
||
import (
|
||
"context"
|
||
"errors"
|
||
|
||
"github.com/cloudwego/eino-ext/components/model/ark"
|
||
"github.com/cloudwego/eino/compose"
|
||
"github.com/cloudwego/eino/schema"
|
||
)
|
||
|
||
const (
|
||
// 图节点：意图识别与约束提取
|
||
schedulePlanGraphNodePlan = "schedule_plan_plan"
|
||
// 图节点：调用粗排算法生成候选方案
|
||
schedulePlanGraphNodePreview = "schedule_plan_preview"
|
||
// 图节点：将候选方案转换为可落库结构
|
||
schedulePlanGraphNodeMaterialize = "schedule_plan_materialize"
|
||
// 图节点：执行落库
|
||
schedulePlanGraphNodeApply = "schedule_plan_apply"
|
||
// 图节点：分析失败原因并生成修补方案
|
||
schedulePlanGraphNodeReflect = "schedule_plan_reflect"
|
||
// 图节点：生成最终回复文案
|
||
schedulePlanGraphNodeFinalize = "schedule_plan_finalize"
|
||
// 图节点：退出（用于提前终止分支）
|
||
schedulePlanGraphNodeExit = "schedule_plan_exit"
|
||
// 图节点：构建混合日程（ReAct 精排前置）
|
||
schedulePlanGraphNodeHybridBuild = "schedule_plan_hybrid_build"
|
||
// 图节点：ReAct 精排循环
|
||
schedulePlanGraphNodeReactRefine = "schedule_plan_react_refine"
|
||
// 图节点：返回精排预览结果（不落库）
|
||
schedulePlanGraphNodeReturnPreview = "schedule_plan_return_preview"
|
||
)
|
||
|
||
// SchedulePlanGraphRunInput 是运行"智能排程 graph"所需的输入依赖。
|
||
//
|
||
// 说明：
|
||
// 1) EmitStage 可选，用于把节点进度推送给外层（例如 SSE 状态块）；
|
||
// 2) Extra 传递前端附加参数（如 task_class_id）；
|
||
// 3) ChatHistory 用于连续对话微调场景。
|
||
type SchedulePlanGraphRunInput struct {
|
||
Model *ark.ChatModel
|
||
State *SchedulePlanState
|
||
Deps SchedulePlanToolDeps
|
||
UserMessage string
|
||
Extra map[string]any
|
||
ChatHistory []*schema.Message
|
||
EmitStage func(stage, detail string)
|
||
// ── ReAct 精排所需 ──
|
||
OutChan chan<- string // SSE 流式输出通道，用于推送 reasoning_content
|
||
ModelName string // 模型名称，用于构造 OpenAI 兼容 chunk
|
||
}
|
||
|
||
// RunSchedulePlanGraph 执行"智能排程"图编排。
|
||
//
|
||
// 图结构：
|
||
//
|
||
// START -> plan -> [branch] -> preview -> [branch] -> materialize -> [branch] -> apply -> [branch]
|
||
// | | | |
|
||
// exit exit exit finalize (成功)
|
||
// |
|
||
// reflect -> [branch] -> apply (重试)
|
||
// |
|
||
// finalize (放弃)
|
||
//
|
||
// 该文件只负责"连线与分支"，节点内部逻辑全部下沉到 nodes.go。
|
||
func RunSchedulePlanGraph(ctx context.Context, input SchedulePlanGraphRunInput) (*SchedulePlanState, error) {
|
||
// 1. 启动前硬校验：模型、状态、依赖缺一不可。
|
||
if input.Model == nil {
|
||
return nil, errors.New("schedule plan graph: model is nil")
|
||
}
|
||
if input.State == nil {
|
||
return nil, errors.New("schedule plan graph: state is nil")
|
||
}
|
||
if err := input.Deps.validate(); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 2. 统一封装阶段推送函数，避免各节点反复判空。
|
||
emitStage := func(stage, detail string) {
|
||
if input.EmitStage != nil {
|
||
input.EmitStage(stage, detail)
|
||
}
|
||
}
|
||
|
||
// 3. 构造 runner，收口节点依赖。
|
||
runner := newSchedulePlanRunner(
|
||
input.Model,
|
||
input.Deps,
|
||
emitStage,
|
||
input.UserMessage,
|
||
input.Extra,
|
||
input.ChatHistory,
|
||
input.OutChan,
|
||
input.ModelName,
|
||
)
|
||
|
||
// 4. 创建状态图容器：输入/输出类型都为 *SchedulePlanState。
|
||
graph := compose.NewGraph[*SchedulePlanState, *SchedulePlanState]()
|
||
|
||
// 5. 注册节点。
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodePlan, compose.InvokableLambda(runner.planNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodePreview, compose.InvokableLambda(runner.previewNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeMaterialize, compose.InvokableLambda(runner.materializeNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeApply, compose.InvokableLambda(runner.applyNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeReflect, compose.InvokableLambda(runner.reflectNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeFinalize, compose.InvokableLambda(runner.finalizeNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeExit, compose.InvokableLambda(runner.exitNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeHybridBuild, compose.InvokableLambda(runner.hybridBuildNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeReactRefine, compose.InvokableLambda(runner.reactRefineNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
if err := graph.AddLambdaNode(schedulePlanGraphNodeReturnPreview, compose.InvokableLambda(runner.returnPreviewNode)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// ── 连线 ──
|
||
|
||
// 6. START -> plan
|
||
if err := graph.AddEdge(compose.START, schedulePlanGraphNodePlan); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 7. plan -> [branch] -> preview | exit
|
||
if err := graph.AddBranch(schedulePlanGraphNodePlan, compose.NewGraphBranch(
|
||
runner.nextAfterPlan,
|
||
map[string]bool{
|
||
schedulePlanGraphNodePreview: true,
|
||
schedulePlanGraphNodeExit: true,
|
||
},
|
||
)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 8. preview -> [branch] -> hybridBuild | materialize | exit
|
||
if err := graph.AddBranch(schedulePlanGraphNodePreview, compose.NewGraphBranch(
|
||
runner.nextAfterPreview,
|
||
map[string]bool{
|
||
schedulePlanGraphNodeHybridBuild: true,
|
||
schedulePlanGraphNodeMaterialize: true,
|
||
schedulePlanGraphNodeExit: true,
|
||
},
|
||
)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 8.1 hybridBuild -> [branch] -> reactRefine | exit
|
||
if err := graph.AddBranch(schedulePlanGraphNodeHybridBuild, compose.NewGraphBranch(
|
||
runner.nextAfterHybridBuild,
|
||
map[string]bool{
|
||
schedulePlanGraphNodeReactRefine: true,
|
||
schedulePlanGraphNodeExit: true,
|
||
},
|
||
)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 8.2 reactRefine -> returnPreview（固定边）
|
||
if err := graph.AddEdge(schedulePlanGraphNodeReactRefine, schedulePlanGraphNodeReturnPreview); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 8.3 returnPreview -> END
|
||
if err := graph.AddEdge(schedulePlanGraphNodeReturnPreview, compose.END); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 9. materialize -> [branch] -> apply | exit
|
||
if err := graph.AddBranch(schedulePlanGraphNodeMaterialize, compose.NewGraphBranch(
|
||
runner.nextAfterMaterialize,
|
||
map[string]bool{
|
||
schedulePlanGraphNodeApply: true,
|
||
schedulePlanGraphNodeExit: true,
|
||
},
|
||
)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 10. apply -> [branch] -> finalize | reflect
|
||
if err := graph.AddBranch(schedulePlanGraphNodeApply, compose.NewGraphBranch(
|
||
runner.nextAfterApply,
|
||
map[string]bool{
|
||
schedulePlanGraphNodeFinalize: true,
|
||
schedulePlanGraphNodeReflect: true,
|
||
},
|
||
)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 11. reflect -> [branch] -> apply (重试) | finalize (放弃)
|
||
if err := graph.AddBranch(schedulePlanGraphNodeReflect, compose.NewGraphBranch(
|
||
runner.nextAfterReflect,
|
||
map[string]bool{
|
||
schedulePlanGraphNodeApply: true,
|
||
schedulePlanGraphNodeFinalize: true,
|
||
},
|
||
)); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 12. finalize -> END
|
||
if err := graph.AddEdge(schedulePlanGraphNodeFinalize, compose.END); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 13. exit -> END
|
||
if err := graph.AddEdge(schedulePlanGraphNodeExit, compose.END); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 14. 运行步数上限：原有链路 ~10 步 + ReAct 精排（hybridBuild + reactRefine + returnPreview = 3）。
|
||
// 加余量到 25，防止异常分支导致无限循环。
|
||
maxSteps := 25
|
||
|
||
// 15. 编译图得到可执行实例。
|
||
runnable, err := graph.Compile(ctx,
|
||
compose.WithGraphName("SchedulePlanGraph"),
|
||
compose.WithMaxRunSteps(maxSteps),
|
||
compose.WithNodeTriggerMode(compose.AnyPredecessor),
|
||
)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 16. 执行图并返回最终状态。
|
||
return runnable.Invoke(ctx, input.State)
|
||
}
|