Version: 0.9.23.dev.260416

后端:
1. Memory 管理面 API 落地(“我的记忆”增删改查 + 恢复)
   - 补齐 List/Get/Create/Update/Delete/Restore 的 handler、请求模型与返回视图
   - 注册 `/api/v1/memory/items*` 路由并接入 MemoryHandler
   - 新增 memory item not found / invalid memory type / invalid memory content 三类管理面错误码
2. Memory Module / Service / Repo 扩展为“可管理 + 可治理”门面
   - 新增 NewModuleWithObserve / ObserveDeps,导出 GetItem / CreateItem / UpdateItem / DeleteItem / RestoreItem / RunDedupCleanup / MemoryObserver / MemoryMetrics
   - 新增手动新增、修改、恢复能力;删除链路切到 SoftDeleteByID;所有管理动作统一事务内写 audit,并桥接向量同步与管理面观测
   - 补齐 CreateItemFields / UpdateItemFields、单条 Create、管理侧字段更新、软删/恢复,以及 dedup 扫描/归档所需 repo 能力
   - 审计操作补齐 archive / restore
3. Memory 读侧与注入侧观测补齐
   - HybridRetrieve 返回 telemetry,统一记录 pinned hit / semantic hit / dedup drop / degraded / RAG fallback,并上报读取命中、去重丢弃、RAG 降级指标
   - AgentService 持有 memory observer / metrics;injectMemoryContext 对读取失败、空注入、成功注入补齐结构化日志与注入计数
4. Worker / 决策 / 向量同步链路治理增强
   - 召回结果显式携带 fallbackMode;hash 精确命中、rag→mysql 降级、最终动作统一写入决策观测
   - 接入 vectorSyncer / observer / metrics;为 job 重试、任务成功/失败、决策分布与 fallback 补齐打点;向量 upsert/delete 统一改走公共 Syncer,并收敛 parseMemoryID 解析逻辑
5. 启动层接入 Memory 观测依赖
   - 启动时创建 LoggerObserver + MetricsRegistry,并通过 NewModuleWithObserve 注入 memory 模块
前端:无
仓库:无
This commit is contained in:
Losita
2026-04-16 19:34:32 +08:00
parent a1b2ffedb8
commit fad3aed30a
23 changed files with 2527 additions and 121 deletions

View File

@@ -10,6 +10,7 @@ import (
infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
memoryobserve "github.com/LoveLosita/smartflow/backend/memory/observe"
memoryrepo "github.com/LoveLosita/smartflow/backend/memory/repo"
memoryutils "github.com/LoveLosita/smartflow/backend/memory/utils"
"github.com/LoveLosita/smartflow/backend/model"
@@ -31,6 +32,26 @@ type ReadService struct {
settingsRepo *memoryrepo.SettingsRepo
ragRuntime infrarag.Runtime
cfg memorymodel.Config
observer memoryobserve.Observer
metrics memoryobserve.MetricsRecorder
}
type retrieveTelemetry struct {
ReadMode string
QueryLen int
LegacyHitCount int
PinnedHitCount int
SemanticHitCount int
DedupDropCount int
FinalCount int
Degraded bool
RAGFallbackUsed bool
}
type semanticRetrieveTelemetry struct {
HitCount int
Degraded bool
RAGFallbackUsed bool
}
func NewReadService(
@@ -38,12 +59,22 @@ func NewReadService(
settingsRepo *memoryrepo.SettingsRepo,
ragRuntime infrarag.Runtime,
cfg memorymodel.Config,
observer memoryobserve.Observer,
metrics memoryobserve.MetricsRecorder,
) *ReadService {
if observer == nil {
observer = memoryobserve.NewNopObserver()
}
if metrics == nil {
metrics = memoryobserve.NewNopMetrics()
}
return &ReadService{
itemRepo: itemRepo,
settingsRepo: settingsRepo,
ragRuntime: ragRuntime,
cfg: cfg,
observer: observer,
metrics: metrics,
}
}
@@ -60,9 +91,14 @@ func (s *ReadService) Retrieve(ctx context.Context, req memorymodel.RetrieveRequ
if now.IsZero() {
now = time.Now()
}
telemetry := retrieveTelemetry{
ReadMode: s.cfg.EffectiveReadMode(),
QueryLen: len(strings.TrimSpace(req.Query)),
}
setting, err := s.settingsRepo.GetByUserID(ctx, req.UserID)
if err != nil {
s.recordRetrieve(ctx, req, telemetry, err)
return nil, err
}
effectiveSetting := memoryutils.EffectiveUserSetting(setting, req.UserID)
@@ -72,16 +108,29 @@ func (s *ReadService) Retrieve(ctx context.Context, req memorymodel.RetrieveRequ
limit := normalizeLimit(req.Limit, defaultRetrieveLimit, maxRetrieveLimit)
if s.cfg.EffectiveReadMode() == memorymodel.MemoryReadModeHybrid {
return s.HybridRetrieve(ctx, req, effectiveSetting, limit, now)
items, hybridTelemetry, hybridErr := s.HybridRetrieve(ctx, req, effectiveSetting, limit, now)
hybridTelemetry.ReadMode = memorymodel.MemoryReadModeHybrid
hybridTelemetry.QueryLen = telemetry.QueryLen
s.recordRetrieve(ctx, req, hybridTelemetry, hybridErr)
return items, hybridErr
}
if s.cfg.RAGEnabled && s.ragRuntime != nil && strings.TrimSpace(req.Query) != "" {
items, ragErr := s.retrieveByRAG(ctx, req, effectiveSetting, limit, now)
if ragErr == nil && len(items) > 0 {
telemetry.SemanticHitCount = len(items)
telemetry.FinalCount = len(items)
s.recordRetrieve(ctx, req, telemetry, nil)
return items, nil
}
telemetry.Degraded = true
telemetry.RAGFallbackUsed = true
}
return s.retrieveByLegacy(ctx, req, limit, now, effectiveSetting)
items, legacyErr := s.retrieveByLegacy(ctx, req, limit, now, effectiveSetting)
telemetry.LegacyHitCount = len(items)
telemetry.FinalCount = len(items)
s.recordRetrieve(ctx, req, telemetry, legacyErr)
return items, legacyErr
}
func (s *ReadService) retrieveByLegacy(
@@ -180,6 +229,58 @@ func normalizeRetrieveMemoryTypes(raw []string) []string {
}
}
func (s *ReadService) recordRetrieve(
ctx context.Context,
req memorymodel.RetrieveRequest,
telemetry retrieveTelemetry,
err error,
) {
if s == nil {
return
}
level := memoryobserve.LevelInfo
if err != nil {
level = memoryobserve.LevelWarn
}
s.observer.Observe(ctx, memoryobserve.Event{
Level: level,
Component: memoryobserve.ComponentRead,
Operation: memoryobserve.OperationRetrieve,
Fields: map[string]any{
"user_id": req.UserID,
"read_mode": telemetry.ReadMode,
"query_len": telemetry.QueryLen,
"legacy_hit_count": telemetry.LegacyHitCount,
"pinned_hit_count": telemetry.PinnedHitCount,
"semantic_hit_count": telemetry.SemanticHitCount,
"dedup_drop_count": telemetry.DedupDropCount,
"final_count": telemetry.FinalCount,
"degraded": telemetry.Degraded,
"rag_fallback_used": telemetry.RAGFallbackUsed,
"success": err == nil,
"error": err,
"error_code": memoryobserve.ClassifyError(err),
},
})
if telemetry.FinalCount > 0 {
s.metrics.AddCounter(memoryobserve.MetricRetrieveHitTotal, int64(telemetry.FinalCount), map[string]string{
"read_mode": strings.TrimSpace(telemetry.ReadMode),
})
}
if telemetry.DedupDropCount > 0 {
s.metrics.AddCounter(memoryobserve.MetricRetrieveDedupDropTotal, int64(telemetry.DedupDropCount), map[string]string{
"read_mode": strings.TrimSpace(telemetry.ReadMode),
})
}
if telemetry.RAGFallbackUsed {
s.metrics.AddCounter(memoryobserve.MetricRAGFallbackTotal, 1, map[string]string{
"read_mode": strings.TrimSpace(telemetry.ReadMode),
})
}
}
// scoreRetrievedItem 计算 legacy 读链路的确定性排序分数。
//
// 说明: