Files
ai-agent/workflow/service/flow/lambda_node.go
T
19904408334 b3f4b94b21 feat: 支持模板子流程递归拷贝与段级续跑优化
* 模板拷贝用户流程时递归复制子流程,并重写 sub_flow 节点引用
* 段级续跑改用列表位置作为段身份,移除对 segment_index 的依赖
* 段结果保存移到每段生成完成时立即落库,降低崩溃丢失风险
* 移除视频分段续跑设计与对应测试
2026-08-25 13:56:55 +08:00

743 lines
27 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
package flow
import (
"ai-agent/gateway"
"ai-agent/workflow/consts/flow"
"ai-agent/workflow/consts/model"
"ai-agent/workflow/consts/node"
"ai-agent/workflow/consts/public"
flowDao "ai-agent/workflow/dao/flow"
nodeDao "ai-agent/workflow/dao/node"
sessionDao "ai-agent/workflow/dao/session"
flowDto "ai-agent/workflow/model/dto/flow"
nodeDto "ai-agent/workflow/model/dto/node"
sessionDto "ai-agent/workflow/model/dto/session"
"ai-agent/workflow/model/entity"
"ai-agent/workflow/service/flow/processor"
"ai-agent/workflow/service/flow/processor/builtin/media"
"context"
"encoding/base64"
"encoding/json"
"fmt"
"strings"
"sync"
"gitea.redpowerfuture.com/red-future/common/db/gfdb"
"gitea.redpowerfuture.com/red-future/common/utils"
"github.com/cloudwego/eino-examples/compose/batch/batch"
"github.com/cloudwego/eino/compose"
"github.com/gogf/gf/v2/database/gdb"
"github.com/gogf/gf/v2/frame/g"
"github.com/gogf/gf/v2/util/gconv"
"github.com/google/uuid"
)
// StartLambda 启动节点
func StartLambda(ctx context.Context, input any) (any, error) {
return input, nil
}
// FormLambda 表单调用节点
func FormLambda(ctx context.Context, input any) (any, error) {
nodeInput, ok := input.(*flowDto.NodeExecutionInput)
if !ok {
return nil, fmt.Errorf("入参类型错误")
}
// 解析 valueSource 引用,填充表单节点输出配置(供下游引用)
for _, output := range nodeInput.Config.OutputConfig {
ProcessValueSourceRecursive(output, nodeInput.Global)
}
return nodeInput, nil
}
// ModelLambda 模型调用节点
func ModelLambda(ctx context.Context, input any) (any, error) {
nodeInput, ok := input.(*flowDto.NodeExecutionInput)
if !ok {
return nil, fmt.Errorf("入参类型错误")
}
modelParams, err := BuildModelRequestBody(nodeInput.Config.ModelConfig.ModelRequestParamsPath, nodeInput.Global)
if err != nil {
return nil, err
}
// 2. 前置工具:决定模型调用入参(单次/多次)
// 入参统一为扁平模型请求体(BuildModelRequestBody 输出,key 为点分路径)。
// 分批处理器按默认上限拆分集合字段,其余前置工具(如 split_shots_pipeline)读取扁平参数。
preToolParams := modelParams
paramsList, err := invokePreTool(ctx, nodeInput.Config.PreTool, preToolParams)
if err != nil {
return nil, err
}
// 3. 逐批调用模型,汇总输出(保持请求顺序),累计 token/费用供节点记录落库
var outputRes []map[string]any
var totalTokens int64
var totalCost float64
if len(paramsList) > 1 {
// 段级续跑仅在"多段 + 视频模型"启用;非视频分段(批量文本等)走原逻辑零影响。
// 段身份 = 列表位置(0-based):paramsList 顺序即段序,concat 按列表顺序拼接;
// 位置互不重复且跨 reExecute 稳定(参数一致 → 段数/顺序不变)。不依赖 params 里的
// segment_index——真实链路(上游 split_shots_pipeline 转写 → 下游 split_segment 按
// __segment_fields 拆分,invokePreTool 剥离 __ 内部键)下 paramsList 只有模型参数。
// segVideo=false 时走既有非段级合并路径(全量生成、不落库、不复用),全新执行行为不变,
// 后续自动 concat 判断(独立的 isVideoModel 调用)仍正常执行。
segVideo := isVideoModel(ctx, nodeInput.Config.ModelConfig.ModelId)
// 续跑(!ForceNewRun)时读取该节点已成功段;全新执行不查(BuildExecution 已清旧段),saved 为 nil → 全量重生成
var saved map[int]entity.SegmentRef
if !nodeInput.Global.ForceNewRun && segVideo {
saved, err = flowDao.FlowSegmentResultDao.ListByNode(ctx, nodeInput.Global.ExecutionId, nodeInput.Config.Id)
if err != nil {
return nil, err
}
}
idxList, needGen := planSegmentResume(paramsList, saved)
results := make([][]map[string]any, len(paramsList))
tokenRes := make([]*gateway.ModelCallRes, len(paramsList))
errs := make([]error, len(paramsList))
saveErrs := make([]error, len(paramsList))
isInference := make([]bool, len(paramsList))
var wg sync.WaitGroup
for i, params := range paramsList {
if !needGen[i] {
continue
}
wg.Add(1)
go func(i int, params map[string]any) {
defer wg.Done()
// 视频段每段失败自动重试 1 次(共 2 次尝试);非视频保持单次调用
for attempt := 0; attempt < segmentGenerateMaxAttempts; attempt++ {
results[i], tokenRes[i], isInference[i], errs[i] = ModelCallResultLambda(ctx, nodeInput.Config.ModelConfig.ModelId, nodeInput.Global.SessionId, params, nodeInput.Config.Prompt)
if errs[i] == nil || !segVideo {
break
}
}
// 每段成功立即落库:该段刚成功即持久化,其他段仍在跑/重试时已成功段也不丢;
// 后续段失败或进程崩溃(panic/OOM/kill)时,已完成段已在库中,reExecute 可直接复用
if segVideo && errs[i] == nil {
for _, rec := range results[i] {
key := media.FindVideoKey(rec)
url := media.FindVideoURL(ctx, rec)
if key == "" || url == "" {
continue
}
if err := flowDao.FlowSegmentResultDao.Save(ctx, nodeInput.Global.ExecutionId, nodeInput.Config.Id, idxList[i], key, url); err != nil {
saveErrs[i] = err
}
}
}
}(i, params)
}
wg.Wait()
// 仍有失败段或落库失败 → 节点失败(成功段已立即落库,供下次 reExecute 复用)
for i := range results {
if saveErrs[i] != nil {
return nil, saveErrs[i]
}
if needGen[i] && errs[i] != nil {
return nil, errs[i]
}
if needGen[i] && tokenRes[i] != nil {
totalTokens += tokenRes[i].TotalTokens
totalCost += tokenRes[i].Cost
}
}
if segVideo {
// 复用段 + 新生段按段序号升序合并,concat 按列表顺序拼接 → 顺序保证
outputRes = mergeSegmentOutputs(idxList, needGen, results, saved)
} else {
if isInference[0] {
outputRes = mergeInferenceBatchResults(results)
} else {
for _, res := range results {
outputRes = append(outputRes, res...)
}
}
}
} else {
for _, params := range paramsList {
res, modelRes, _, err := ModelCallResultLambda(ctx, nodeInput.Config.ModelConfig.ModelId, nodeInput.Global.SessionId, params, nodeInput.Config.Prompt)
if err != nil {
return nil, err
}
if modelRes != nil {
totalTokens += modelRes.TotalTokens
totalCost += modelRes.Cost
}
outputRes = append(outputRes, res...)
}
}
// 3.5 把本次节点消耗的 token/费用写入节点执行记录,供汇总节点聚合到 exec_workflow
if nodeInput.NodeExecutionId > 0 && (totalTokens > 0 || totalCost > 0) {
if _, err = nodeDao.NodeExecutionDao.Update(ctx, &nodeDto.UpdateNodeExecutionReq{
Id: nodeInput.NodeExecutionId,
TokenInfo: []map[string]any{{
"total_tokens": totalTokens,
"total_fee": totalCost,
}},
}); err != nil {
return nil, fmt.Errorf("节点:%v 写入token信息失败: %v", nodeInput.Config.Name, err)
}
}
// 4.5 视频模型节点返回多个视频时,自动调用视频合成工具(concat_videos)合并为单条;
// 已显式配置 concat_videos 后置工具时跳过,避免重复合并
if nodeInput.Config.PostTool != media.ProcessorName && len(outputRes) > 1 && isVideoModel(ctx, nodeInput.Config.ModelConfig.ModelId) {
g.Log().Debugf(ctx, "modelId1:%v ,outputRes: %v", nodeInput.Config.ModelConfig.ModelId, outputRes)
outputRes, err = invokePostTool(ctx, media.ProcessorName, outputRes, map[string]any{"callback_url": "callback_url", "upload": true})
g.Log().Debugf(ctx, "modelId2:%v ,outputRes: %v", nodeInput.Config.ModelConfig.ModelId, outputRes)
if err != nil {
return nil, err
}
} else {
// 4. 后置工具:加工模型输出(透传原始请求参数,供后置工具读取合并配置等)
outputRes, err = invokePostTool(ctx, nodeInput.Config.PostTool, outputRes, modelParams)
if err != nil {
return nil, err
}
}
g.Log().Debugf(ctx, "modelId3:%v ,outputRes: %v", nodeInput.Config.ModelConfig.ModelId, outputRes)
nodeInput.Config.OutputResult = outputRes
return nodeInput, nil
}
// isVideoModel 判断模型是否为视频模型(模型类型 TypeVideo=600),用于视频节点多视频自动合成判断
func isVideoModel(ctx context.Context, modelId int64) bool {
modelInfo, err := gateway.GetModelInfoById(ctx, &gateway.GetModelInfoByIdReq{ModelId: modelId})
if err != nil {
g.Log().Warningf(ctx, "查询模型配置失败,跳过自动视频合成 modelId=%d err=%v", modelId, err)
return false
}
return modelInfo.ModelManage.ModelType != nil && *modelInfo.ModelManage.ModelType == model.TypeVideo
}
// mergeInferenceBatchResults 推理模型分批结果拼接为单条输出记录:
// 各批结果按批序对同名 key 的值做字符串拼接("拼到一个字段"),最终返回单条 {key:值} 记录。
// 非字符串值(如结构/数组字段)取最后一份,避免误拼接。
func mergeInferenceBatchResults(results [][]map[string]any) []map[string]any {
merged := make(map[string]any)
for _, res := range results {
for _, record := range res {
for key, val := range record {
prev, has := merged[key]
if !has {
merged[key] = val
continue
}
sPrev, pOK := prev.(string)
sVal, vOK := val.(string)
if pOK && vOK {
merged[key] = sPrev + "\n" + sVal
continue
}
merged[key] = val
}
}
}
return []map[string]any{merged}
}
// invokePreTool 执行前置处理器,把模型请求参数转换为模型调用入参列表。
// 前置处理器契约:入参即模型请求参数本体;返回值:
// - map[string]any 一次模型调用,入参为返回值
// - []map[string]any 多次模型调用,逐个入参请求
// - nil 视为异常,节点失败(不允许静默跳过模型调用)
func invokePreTool(ctx context.Context, processorName string, modelParams map[string]any) (paramsList []map[string]any, err error) {
if processorName == "" {
return []map[string]any{stripInternalKeys(modelParams)}, nil
}
data, err := processor.Call(ctx, processorName, modelParams)
if err != nil {
return nil, fmt.Errorf("执行前置处理器[%s]失败: %v", processorName, err)
}
switch v := data.(type) {
case nil:
return nil, fmt.Errorf("前置处理器[%s]返回空", processorName)
case map[string]any:
return []map[string]any{stripInternalKeys(v)}, nil
case []map[string]any:
list := make([]map[string]any, 0, len(v))
for _, m := range v {
list = append(list, stripInternalKeys(m))
}
return list, nil
default:
return nil, fmt.Errorf("前置处理器[%s]返回类型不支持: %T", processorName, data)
}
}
// stripInternalKeys 剥离 __ 前缀的内部键(如 __segment_fields/__produced),
// 模型网关做参数严格校验(CheckParams strictUnknown)会拒绝未知字段,内部标记不得随请求体下发。
func stripInternalKeys(params map[string]any) map[string]any {
if params == nil {
return params
}
for k := range params {
if strings.HasPrefix(k, "__") {
delete(params, k)
}
}
return params
}
// invokePostTool 执行后置处理器,加工模型调用结果。
// 后置处理器契约:入参 {"output": 模型输出结果列表, "request": 原始模型请求参数}(列表须包成对象传入);返回值:
// - []map[string]any 替换模型输出
// - map[string]any 替换为单条输出
// - nil 保留原输出
func invokePostTool(ctx context.Context, processorName string, outputRes []map[string]any, requestParams map[string]any) ([]map[string]any, error) {
if processorName == "" {
return outputRes, nil
}
data, err := processor.Call(ctx, processorName, map[string]any{"output": outputRes, "request": requestParams})
if err != nil {
return nil, fmt.Errorf("执行后置处理器[%s]失败: %v", processorName, err)
}
switch v := data.(type) {
case nil:
return outputRes, nil
case []map[string]any:
return v, nil
case map[string]any:
return []map[string]any{v}, nil
default:
return nil, fmt.Errorf("后置处理器[%s]返回类型不支持: %T", processorName, data)
}
}
func SubFlowLambda(ctx context.Context, input any) (any, error) {
// 1. 类型断言(和其他节点保持一致的入参结构)
nodeExecInput, ok := input.(*flowDto.NodeExecutionInput)
if !ok {
return nil, fmt.Errorf("子流程节点入参类型错误,期望*flowDto.NodeExecutionInput,实际%T", input)
}
// 2. 解析子流程配置
subFlowConfig := nodeExecInput.Config.SubConfig
if subFlowConfig == nil {
return nil, fmt.Errorf("子流程节点缺少配置")
}
getRes, err := FlowUserService.Get(ctx, &flowDto.GetFlowUserReq{
Id: subFlowConfig.WorkflowId,
})
if err != nil {
return nil, err
}
// 3. 引入参数解析:把首页表单值/上游引用值/静态默认值写入子流程开始节点 outputConfig。
// 须在 BuildGraph / ExtractFlowNodeFrom 之前执行,batchInputs 深拷贝的才是注入后的开始节点。
injectSubFlowFields(nodeExecInput.Global, getRes.FlowContent, subFlowConfig.Fields)
// 4. 并发数:从主流程开始节点 outputConfig 的 maxConcurrency 字段读取(前端把子流程节点生成次数表单字段聚合到主流程开始节点),读不到再用子流程节点配置兜底
maxConcurrency := mainFlowMaxConcurrency(nodeExecInput.Global, subFlowConfig.MaxConcurrency)
// 4. 编译子流程Graph(复用现有 BuildGraphFromFlowContent 逻辑)
nodeList, subGraph := BuildGraph(ctx, getRes.FlowContent)
// 4. 构建子流程Workflow(绑定START/END,和示例对齐)
innerWorkflow := compose.NewWorkflow[*flowDto.FlowExecutionInput, *flowDto.FlowExecutionInput]()
// 挂载子图节点并绑定全局START
innerWorkflow.AddGraphNode("sub_flow_graph", subGraph).AddInput(compose.START)
// 绑定子图输出到全局END
innerWorkflow.End().AddInput("sub_flow_graph")
// 生成次数(批量条数):maxConcurrency<=0 时按 1 次兜底
batchCount := maxConcurrency
if batchCount <= 0 {
batchCount = 1
}
// 5. 构建BatchNode(批量执行子流程,复用示例逻辑)
batchNode := batch.NewBatchNode(&batch.NodeConfig[*flowDto.FlowExecutionInput, *flowDto.FlowExecutionInput]{
Name: fmt.Sprintf("sub_flow_batch_%s", nodeExecInput.Config.Id),
InnerTask: innerWorkflow,
MaxConcurrency: batchCount,
})
// 6. 提取批量输入:按生成次数生成 N 份(每份独立克隆 ConfigMap,避免并发执行时节点输出写串)
nodeInputParams := ExtractFlowNodeFrom(getRes.FlowContent)
configMap := make(map[string]*entity.FlowNode)
for _, cfg := range nodeInputParams {
configMap[cfg.Id] = cfg
}
for _, i := range nodeList {
configMap[i.Id] = &i
}
batchInputs := make([]*flowDto.FlowExecutionInput, 0, batchCount)
for j := 0; j < batchCount; j++ {
batchInputs = append(batchInputs, &flowDto.FlowExecutionInput{
NodeGroupId: nodeExecInput.Global.NodeGroupId,
ExecutionId: nodeExecInput.Global.ExecutionId,
FlowId: nodeExecInput.Global.FlowId,
ConfigMap: cloneConfigMap(configMap),
SessionId: nodeExecInput.Global.SessionId,
})
}
// 7. 执行批量子流程
batchOutput, err := batchNode.Invoke(ctx, batchInputs)
if err != nil {
return nil, fmt.Errorf("执行子流程BatchNode失败: %v", err)
}
// 8. 展平每份子流程执行的节点输出,写回当前节点 OutputResult 供下游引用
var outputRes []map[string]any
for _, single := range batchOutput {
if single == nil {
continue
}
outputRes = append(outputRes, collectFlowNodeResults(single)...)
}
g.Log().Info(ctx, fmt.Sprintf("子流程执行完成,共 %d 次,输出 %d 条", batchCount, len(outputRes)))
nodeExecInput.Config.OutputResult = outputRes
return nodeExecInput, nil
}
// injectSubFlowFields 将子流程节点引入参数(subConfig.Fields)解析后写入子流程开始节点
// outputConfig,使子流程启动时能读到首页表单值/上游引用值/静态默认值。
// 每个字段的取值优先级:valueSource 引用解析成功 → field.value → field.defaultValue
// 匹配键为 field(前端约定以 field 为主,不兼容 path)。
func injectSubFlowFields(global *flowDto.FlowExecutionInput, subFlowContent *entity.FlowInfo, fields []map[string]any) {
if global == nil || subFlowContent == nil || len(fields) == 0 {
return
}
startNode := subFlowStartNode(subFlowContent)
if startNode == nil {
return
}
byField := make(map[string]map[string]any, len(startNode.OutputConfig))
for _, output := range startNode.OutputConfig {
byField[gconv.String(output["field"])] = output
}
for _, field := range fields {
entry := byField[gconv.String(field["field"])]
if entry == nil {
continue
}
value := field["value"]
if vs, has := field["valueSource"]; has && vs != nil {
if vsNodeId, vsField := firstValueSource(vs); vsNodeId != "" && vsField != "" {
if v, _, ok := resolveValueSource(global, vsNodeId, vsField); ok {
value = v
}
}
}
if value == nil {
value = field["defaultValue"]
}
if value != nil {
entry["value"] = value
}
}
}
// firstValueSource 从 valueSource 提取第一个引用源 (nodeId, field)。
// 前端统一发送数组 [{nodeId, field}](见 serializeSubFlowConfig),旧 DSL 可能是单对象
// {nodeId, fieldName},两种形态都兼容;子流程字段与引用源一一对应,只取第一个。
func firstValueSource(vs any) (nodeId, field string) {
if vs == nil {
return
}
switch v := vs.(type) {
case []any:
if len(v) > 0 {
return firstValueSource(v[0])
}
return
case []map[string]any:
if len(v) > 0 {
return firstValueSource(v[0])
}
return
}
m := gconv.Map(vs)
nodeId = gconv.String(m["nodeId"])
field = gconv.String(m["fieldName"])
if field == "" {
field = gconv.String(m["field"])
}
return
}
// subFlowStartNode 返回工作流开始节点
func subFlowStartNode(content *entity.FlowInfo) *entity.FlowNode {
if content == nil {
return nil
}
for i := range content.Nodes {
if content.Nodes[i].Id == content.StartNodeId {
return &content.Nodes[i]
}
}
return nil
}
// mainFlowMaxConcurrency 取子流程批量执行并发数:从主流程开始节点 outputConfig
// 的 maxConcurrency 字段读取(前端把子流程节点的生成次数表单字段聚合到主流程开始节点),
// 读不到再用子流程节点配置的兜底值。
func mainFlowMaxConcurrency(global *flowDto.FlowExecutionInput, fallback int) int {
if global == nil {
return fallback
}
for _, n := range global.ConfigMap {
if n == nil || n.NodeCode != node.NodeTypeStart {
continue
}
for _, output := range n.OutputConfig {
if gconv.String(output["field"]) != "maxConcurrency" && gconv.String(output["path"]) != "maxConcurrency" {
continue
}
if v := gconv.Int(output["value"]); v > 0 {
return v
}
}
return fallback
}
return fallback
}
// cloneConfigMap 深拷贝 ConfigMap,保证各批次子流程并发执行时节点输出互不串扰。
// 浅拷贝会共享 *entity.FlowNode,并发写 OutputResult 产生竞态。
func cloneConfigMap(src map[string]*entity.FlowNode) map[string]*entity.FlowNode {
dst := make(map[string]*entity.FlowNode, len(src))
for k, v := range src {
data, err := json.Marshal(v)
if err != nil {
dst[k] = v
continue
}
n := new(entity.FlowNode)
if err = json.Unmarshal(data, n); err != nil {
dst[k] = v
continue
}
dst[k] = n
}
return dst
}
// collectFlowNodeResults 收集一次子流程执行中所有已执行节点的输出,展平成 {字段:值} 列表
func collectFlowNodeResults(execInput *flowDto.FlowExecutionInput) []map[string]any {
var res []map[string]any
for _, executed := range execInput.ExecutedNodes {
if nodeConfig := execInput.ConfigMap[executed.NodeId]; nodeConfig != nil {
res = append(res, nodeConfig.OutputResult...)
}
}
return res
}
// HttpLambda 构建HTTP(S)接口
func HttpLambda(ctx context.Context, input any) (any, error) {
nodeInput, ok := input.(*flowDto.NodeExecutionInput)
if !ok {
return nil, fmt.Errorf("入参类型错误")
}
outputRes, err := HttpCallResultLambda(ctx, nodeInput)
if err != nil {
return nil, err
}
nodeInput.Config.OutputResult = outputRes
return nodeInput, nil
}
func DataMergeLambda(ctx context.Context, input any) (res any, err error) {
nodeInput, ok := input.(*flowDto.NodeExecutionInput)
if !ok {
return nil, fmt.Errorf("参数合并入参类型错误")
}
return nodeInput, nil
}
func SummaryLambda(ctx context.Context, input any) (any, error) {
execInput, ok := input.(*flowDto.NodeExecutionInput)
if !ok {
return nil, fmt.Errorf("汇总节点入参类型错误,实际是 %T", input)
}
// 聚合所有已执行节点中需入库的文件结果(两层规则)
summaryResult := collectSaveFileResults(ctx, execInput.Global)
// 把汇总结果存入当前节点的输出
g.Log().Info(ctx, fmt.Sprintf("结果汇总完成,汇总数据:%+v", summaryResult))
err := gfdb.DB(ctx, public.DbNameBlackDeacon).Transaction(ctx, func(ctx context.Context, tx gdb.TX) error {
res, _, err := nodeDao.NodeExecutionDao.ListByFlowExecutionId(ctx, &nodeDto.ListNodeExecutionByFlowReq{
NodeGroupId: execInput.Global.NodeGroupId,
}, entity.NodeExecutionCol.TokenInfo)
if err != nil {
return err
}
var totalTokens int
var totalFee float64
for _, item := range res {
for _, itemToken := range item.TokenInfo {
m := gconv.Map(itemToken)
totalTokens += gconv.Int(m["total_tokens"])
totalFee += gconv.Float64(m["total_fee"])
}
}
_, err = sessionDao.ExecWorkflowDao.Update(ctx, &sessionDto.UpdateWorkflowReq{
Id: execInput.Global.ExecutionId,
Status: flow.FlowExecutionStatusSuccess.Code(),
TotalTokens: totalTokens,
TotalFee: totalFee,
})
if err != nil {
return err
}
if len(summaryResult) > 0 {
_, err = sessionDao.ExecWorkflowResultDao.BatchInsert(ctx, summaryResult)
if err != nil {
return err
}
}
return nil
})
return execInput, err
}
// collectSaveFileResults 按两层规则收集需入库的文件结果:
// 第一层:节点须开启"保存文件"IsSaveFile);
// 第二层:key 取自节点 OutputResult 的各字段,命中 ModelResponseBodyMapping 才入库;
// 原始响应体 key(respBody)恒入库(不要求映射声明);HTTP 节点产出以 http_file_url:{key}
// 标记的字段(IsSaveFile 时由 HttpCallResultLambda 生成)恒入库(无模型响应映射可查)。
// 结果值为 http(s) URL 或 MinIO 对象裸路径直接使用;非路径值(base64 图片/文本)先上传 OSS 换取 URL,
// 文本内容以 .inc 扩展名存储。
func collectSaveFileResults(ctx context.Context, execInput *flowDto.FlowExecutionInput) []*sessionDto.CreateWorkflowResultReq {
if execInput == nil {
return nil
}
var summaryResult []*sessionDto.CreateWorkflowResultReq
for _, executedNode := range execInput.ExecutedNodes {
nodeConfig := execInput.ConfigMap[executedNode.NodeId]
if nodeConfig == nil || len(nodeConfig.OutputResult) == 0 || !nodeConfig.IsSaveFile {
continue
}
// 第二层:key 取自节点 OutputResult 的各字段,
// 命中 ModelResponseBodyMapping 才入库;respBody 与 HTTP 节点 http_file_url:{key} 标记恒入库
saveKeys := nodeConfig.ModelConfig.ModelResponseBodyMapping
for _, respBody := range nodeConfig.OutputResult {
for key, val := range gconv.Map(respBody) {
isHTTPFile := strings.HasPrefix(key, "http_file_url:")
if !isHTTPFile {
if _, ok := saveKeys[key]; !ok && key != "respBody" {
continue
}
}
fileUrl, err := resolveSaveFileResult(ctx, val)
if err != nil {
g.Log().Warningf(ctx, "collectSaveFileResults 上传结果文件失败 key=%s err=%v", key, err)
continue
}
summaryResult = append(summaryResult, &sessionDto.CreateWorkflowResultReq{
SessionId: execInput.SessionId,
FlowId: execInput.FlowId,
ExecId: execInput.ExecutionId,
ResultFileUrl: fileUrl,
})
}
}
}
return summaryResult
}
// resolveSaveFileResult 解析结果值为可入库的 URL:
// - 已是 http(s) URL 或 MinIO 对象裸路径 → 直接返回
// - 非路径(base64 图片/文本)→ 上传 OSS 换取 URL
func resolveSaveFileResult(ctx context.Context, val any) (string, error) {
isPath, path, fileBytes, ext := resolveFileContent(val)
if isPath {
return path, nil
}
if ext == "" {
ext = ".png"
}
fileUrl, err := gateway.Upload(ctx, fmt.Sprintf("workflow_result_%s%s", uuid.NewString(), ext), fileBytes)
if err != nil {
return "", err
}
return fileUrl, nil
}
// resolveFileContent 判断结果值形态:
// - 已是 URL 路径(http/https 开头)→ 直接使用
// - data URIdata:<mime>;base64,<data>)→ 解码为字节,扩展名按 mime 推断
// - 纯 base64(可解码且长度足以认为是编码数据)→ 解码为字节,默认 .png
// - 其余(文本)→ 以 .inc 扩展名上传原文
func resolveFileContent(val any) (isPath bool, path string, fileBytes []byte, ext string) {
s := gconv.String(val)
if isFileURL(s) {
return true, s, nil, ""
}
// MinIO 对象裸路径(无 http 前缀,模型网关转存 OSS 后返回)
if utils.IsOSSPath(s) {
return true, s, nil, ""
}
// data URIdata:<mime>;base64,<payload>
if b, mime, ok := parseDataURI(s); ok {
return false, "", b, extOfMime(mime)
}
// 纯 base64:可解码且长度足够,视为编码后的文件内容
trimmed := strings.TrimSpace(s)
if len(trimmed) >= 64 {
if b, err := base64.StdEncoding.DecodeString(trimmed); err == nil && len(b) > 0 {
return false, "", b, ".png"
}
}
// 文本:以 .inc 存储
return false, "", []byte(s), ".inc"
}
// isFileURL 判断字符串是否已是对外可访问的 URL 路径(http/https 开头)
func isFileURL(s string) bool {
lower := strings.ToLower(s)
return strings.HasPrefix(lower, "http://") || strings.HasPrefix(lower, "https://")
}
// extOfMime 按 MIME 类型推断文件扩展名
func extOfMime(mime string) string {
switch strings.ToLower(strings.TrimSpace(mime)) {
case "image/png", "png":
return ".png"
case "image/jpeg", "image/jpg", "jpeg", "jpg":
return ".jpg"
case "image/webp":
return ".webp"
case "image/gif":
return ".gif"
case "audio/mpeg", "audio/mp3", "mp3":
return ".mp3"
case "audio/wav", "wav":
return ".wav"
case "video/mp4", "mp4":
return ".mp4"
case "application/json", "json":
return ".json"
default:
return ""
}
}
// parseDataURI 解析 data URIdata:<mime>;base64,<payload>,返回解码字节与 mime
func parseDataURI(s string) ([]byte, string, bool) {
const prefix = "data:"
if !strings.HasPrefix(s, prefix) {
return nil, "", false
}
rest := s[len(prefix):]
comma := strings.Index(rest, ",")
if comma < 0 {
return nil, "", false
}
mime := rest[:comma]
if semicolon := strings.Index(mime, ";"); semicolon >= 0 {
mime = mime[:semicolon]
}
payload := strings.TrimPrefix(rest[comma+1:], "base64,")
b, err := base64.StdEncoding.DecodeString(payload)
if err != nil {
return nil, "", false
}
return b, mime, true
}