diff --git a/go/internal/storage/lancedb_ipc.go b/go/internal/storage/lancedb_ipc.go index 55afe70..ca3036e 100644 --- a/go/internal/storage/lancedb_ipc.go +++ b/go/internal/storage/lancedb_ipc.go @@ -7,7 +7,6 @@ import ( "fmt" "io" "log" - "math" "net" "os" "path/filepath" @@ -257,62 +256,32 @@ func (rc *RustLanceDBClient) Search(table string, vector []float32, topK int, na if namespaceFilter != "" && r.Namespace != namespaceFilter { continue } - // 重要性计算:recency_factor × (1 + log(1 + recall_count)) + // 重要性计算 + 质量融合 —— 纯函数在 quality_fusion.go(t_5028c14d) daysSinceRecall := 30.0 // default for never-recalled if r.LastRecalledAt != "" { if lastT, err := time.Parse(time.RFC3339, r.LastRecalledAt); err == nil { daysSinceRecall = time.Since(lastT).Hours() / 24.0 } } - recencyFactor := 1.0 - if daysSinceRecall > 7 { - recencyFactor = 0.5 // 7天内高权重,之后折半 - } - if daysSinceRecall > 30 { - recencyFactor = 0.2 // 30天以上权重很低 - } // GA-P0-2 (t_c6d41ac6):排序去偏用 recall_seen_count(观测/选择信号); // recall_verified_count 只参与衰减/遗忘决策,不参与排序。 - recallCount := float64(r.RecallSeenCount) - computedImportance := recencyFactor * (1.0 + math.Log(1.0+recallCount)) - - // VolatileFlag 惩罚 - if r.VolatileFlag { - computedImportance *= 0.5 - } + computedImportance := computeImportance(daysSinceRecall, r.RecallSeenCount, r.VolatileFlag) // QualityScore 融合(归一化到 0-1 范围) // t_631ab13d:先存原始 DB 值(r.QualityScore 是 LanceDB 列直接值), // 融合后 finalQuality 写进 QualityScore(供排序),RawQualityScore 保留原值(供 recall 响应展示) dbRawQS := r.QualityScore - finalQuality := math.Min(computedImportance, 1.0) // 2026-10-12:存量 quality_score 有 54.2% 不可信(0=未打分 / 0.755=历史常量 bug)。 // 修入口治不了存量 —— 消费端必须自己判定,不可信时把权重让给可信信号。 - qsTrusted := models.IsReliableQuality(r.QualityScore) - if qsTrusted { - // QualityScore from LLM (0-1) + computed importance, capped at 1.0 - finalQuality = math.Min(0.4*computedImportance+0.6*r.QualityScore, 1.0) - } else if qstrustVerbose { + if !models.IsReliableQuality(r.QualityScore) && qstrustVerbose { log.Printf("[qstrust-ipc] 不可信质量分已剔除 id=%s rawQS=%.6f → 仅用 importance %.4f", r.ID, r.QualityScore, computedImportance) } - // V 值融合:用户决策信号(correct/useful/not-useful) + var vVal float64 if VValueProvider != nil { - vVal := VValueProvider(r.ID) - if vVal > 0 { - if qsTrusted { - // 30% V值 + 30% importance + 40% LLM quality - finalQuality = 0.3*vVal + 0.3*computedImportance + 0.4*r.QualityScore - } else { - // 质量分不可信 → 去掉它的 40%,剩余 30:30 重新归一为 50:50。 - // 这样既不让假分数进来,也不因为"少了一项"而整体压沉。 - finalQuality = 0.5*vVal + 0.5*computedImportance - } - if finalQuality > 1.0 { - finalQuality = 1.0 - } - } + vVal = VValueProvider(r.ID) } + finalQuality := fuseQuality(computedImportance, r.QualityScore, vVal) out = append(out, models.MemoryRecord{ ID: r.ID, Content: r.Content, Category: r.Category, Namespace: r.Namespace, diff --git a/go/internal/storage/quality_fusion.go b/go/internal/storage/quality_fusion.go new file mode 100644 index 0000000..c729dd0 --- /dev/null +++ b/go/internal/storage/quality_fusion.go @@ -0,0 +1,63 @@ +package storage + +import ( + "math" + + "github.com/xiaoxue/memoryweave/internal/models" +) + +// 召回排序「频率 + 时间衰减」融合公式 —— 从 lancedb_ipc.go Search() 内联代码抽出的纯函数。 +// 行为逐位等价于原内联实现(t_5028c14d),不改任何数值常量。 +// 调用链:recall.go Recall() → Search() → 本文件三函数 → QualityScore 供排序。 + +// computeRecencyFactor 时间衰减因子:>30d→0.2 / >7d→0.5 / else 1.0(原 :267-273)。 +func computeRecencyFactor(days float64) float64 { + if days > 30 { + return 0.2 // 30天以上权重很低 + } + if days > 7 { + return 0.5 // 7天内高权重,之后折半 + } + return 1.0 +} + +// computeImportance 重要度 = recencyFactor × (1 + log(1 + recall_seen_count)); +// volatile 减半(原 :260-282)。recall_seen_count 是观测/选择信号(GA-P0-2), +// recall_verified_count 只参与衰减/遗忘决策,不参与排序。 +func computeImportance(days float64, recallSeenCount int64, volatile bool) float64 { + rf := computeRecencyFactor(days) + imp := rf * (1.0 + math.Log(1.0+float64(recallSeenCount))) + if volatile { + imp *= 0.5 + } + return imp +} + +// fuseQuality 融合 QualityScore(归一化 0-1):computedImportance 与 LLM 质量分、 +// 用户决策 V 值按可信度分支加权(原 :284-315)。 +// vValue ≤ 0 表示无用户决策信号(VValueProvider nil 或返回非正)→ 不进 V 分支。 +// rawQuality 经 models.IsReliableQuality 判定可信度(0=未打分/历史常量 bug 视为不可信)。 +func fuseQuality(computedImportance, rawQuality, vValue float64) float64 { + qsTrusted := models.IsReliableQuality(rawQuality) + var finalQuality float64 + if qsTrusted { + // QualityScore from LLM (0-1) + computed importance, capped at 1.0 + finalQuality = math.Min(0.4*computedImportance+0.6*rawQuality, 1.0) + } else { + // 质量分不可信 → 仅用 importance([qstrust-ipc] 日志留在调用处) + finalQuality = math.Min(computedImportance, 1.0) + } + if vValue > 0 { + if qsTrusted { + // 30% V值 + 30% importance + 40% LLM quality + finalQuality = 0.3*vValue + 0.3*computedImportance + 0.4*rawQuality + } else { + // 质量分不可信 → 去掉它的 40%,剩余 30:30 重新归一为 50:50。 + finalQuality = 0.5*vValue + 0.5*computedImportance + } + if finalQuality > 1.0 { + finalQuality = 1.0 + } + } + return finalQuality +} \ No newline at end of file diff --git a/go/internal/storage/quality_fusion_test.go b/go/internal/storage/quality_fusion_test.go new file mode 100644 index 0000000..06e00b0 --- /dev/null +++ b/go/internal/storage/quality_fusion_test.go @@ -0,0 +1,87 @@ +package storage + +import ( + "math" + "testing" +) + +// 锁死「频率/时间衰减参与召回排序」:MOA 三件套 ③ 测试覆盖。 +// 行为逐位等价于原 lancedb_ipc.go Search() 内联公式(AC3 手工期望值)。 + +func TestComputeRecencyFactor(t *testing.T) { + cases := []struct { + days float64 + want float64 + }{ + {1.0, 1.0}, // ≤7d 高权重 + {7.0, 1.0}, // 边界:>7 才折半 + {10.0, 0.5}, // 7~30d 折半 + {30.0, 0.5}, // 边界:>30 才再降 + {31.0, 0.2}, // >30d 很低 + {100.0, 0.2}, + } + for _, c := range cases { + if got := computeRecencyFactor(c.days); got != c.want { + t.Errorf("computeRecencyFactor(%v)=%v, want %v", c.days, got, c.want) + } + } +} + +func TestComputeImportance(t *testing.T) { + // seen=0,days=1 → rf=1, imp=1*(1+log(1))=1.0 + if got := computeImportance(1.0, 0, false); got != 1.0 { + t.Errorf("imp(seen=0,days=1)=%v, want 1.0", got) + } + // seen=7,days=1 → 1*(1+ln8)=3.0794… + if got := computeImportance(1.0, 7, false); math.Abs(got-(1+math.Log(8))) > 1e-9 { + t.Errorf("imp(seen=7,days=1)=%v, want %.4f", got, 1+math.Log(8)) + } + // seen=0,days=10 → rf=0.5, imp=0.5 + if got := computeImportance(10.0, 0, false); got != 0.5 { + t.Errorf("imp(seen=0,days=10)=%v, want 0.5", got) + } + // 同样频次+同样时间,volatile 减半:seen=7,days=1,volatile → 0.5*3.0794 + if got := computeImportance(1.0, 7, true); math.Abs(got-0.5*(1+math.Log(8))) > 1e-9 { + t.Errorf("imp(volatile)=%v, want %.4f", got, 0.5*(1+math.Log(8))) + } + // 频率单调性:同 rf 下 seen 越大 imp 越大 + a := computeImportance(1.0, 0, false) + b := computeImportance(1.0, 3, false) + c := computeImportance(1.0, 7, false) + if !(a < b && b < c) { + t.Errorf("imp 应随频率单调增,got a=%v b=%v c=%v", a, b, c) + } +} + +func TestFuseQuality(t *testing.T) { + // raw 可信 + 无 V 值 → 0.4*imp+0.6*raw, cap 1.0 + // raw=0.605, imp=1.0 → 0.4*1+0.6*0.605=0.763 + if got := fuseQuality(1.0, 0.605, 0); math.Abs(got-0.763) > 1e-9 { + t.Errorf("fuse(imp=1,raw=0.605,可信)=%v, want 0.763", got) + } + // 频率越高 → 融合分越高(可信+无V 分支) + lo := fuseQuality(computeImportance(1.0, 0, false), 0.605, 0) + hi := fuseQuality(computeImportance(1.0, 7, false), 0.605, 0) + if !(lo < hi) { + t.Errorf("fuse 应随 frequency 增,got lo=%v hi=%v", lo, hi) + } + // 越近 → 融合分越高 + old := fuseQuality(computeImportance(31.0, 7, false), 0.605, 0) + near := fuseQuality(computeImportance(1.0, 7, false), 0.605, 0) + if !(old < near) { + t.Errorf("fuse 应随 recency 增,got old=%v near=%v", old, near) + } + // V 值分支:可信 → 0.3*v + 0.3*imp + 0.4*raw + if got := fuseQuality(1.0, 0.605, 0.9); math.Abs(got-(0.3*0.9+0.3*1.0+0.4*0.605)) > 1e-9 { + t.Errorf("fuse(V=0.9,可信)=%v, want %v", got, 0.3*0.9+0.3*1.0+0.4*0.605) + } + // 不可信 + 有 V → 0.5*v+0.5*imp(raw 被剔除) + // raw=0(不可信), imp=1.0, v=0.9 → 0.5*0.9+0.5*1.0=0.95 + if got := fuseQuality(1.0, 0, 0.9); math.Abs(got-0.95) > 1e-9 { + t.Errorf("fuse(V=0.9,不可信)=%v, want 0.95", got) + } + // 融合结果有界 [0,1] + if got := fuseQuality(10.0, 1.0, 1.0); got > 1.0 { + t.Errorf("fuse 应 cap 到 1.0,got %v", got) + } +} \ No newline at end of file