fix distill threshold: allow soulful short facts even when LLM score is low

This commit is contained in:
小唯
2026-09-19 23:51:32 +08:00
parent ab974d0bed
commit 3e8157307e
+34 -23
View File
@@ -74,11 +74,11 @@ type LLMResponse struct {
VD float64 `json:"vd"`
RU float64 `json:"ru"`
Entities []string `json:"entities"`
Facts []string `json:"facts"` // 旧格式兼容
Decisions []string `json:"decisions"` // 新格式:决策结论
Conclusions []string `json:"conclusions"` // 新格式:最终结论
ActionsTaken []string `json:"actions_taken"` // 新格式:采取的行动
OpenQuestions []string `json:"open_questions"` // 新格式:悬而未决
Facts []string `json:"facts"` // 旧格式兼容
Decisions []string `json:"decisions"` // 新格式:决策结论
Conclusions []string `json:"conclusions"` // 新格式:最终结论
ActionsTaken []string `json:"actions_taken"` // 新格式:采取的行动
OpenQuestions []string `json:"open_questions"` // 新格式:悬而未决
}
// LLM 5维权重
@@ -102,19 +102,19 @@ type Engine struct {
APIKey string
// 队列
queue []DistillInput
batchSize int
queue []DistillInput
batchSize int
batchTimeout time.Duration
lastDistill time.Time
lastDistill time.Time
// P3 双缓冲:token 积累触发(LightMem 移植)
pendingTokens int
flushTokenThreshold int
// 成本控制
dailyLimit int
dailyUsed int
dailyReset time.Time
dailyLimit int
dailyUsed int
dailyReset time.Time
// HTTP客户端
client *http.Client
@@ -250,10 +250,15 @@ func (e *Engine) distillOne(input DistillInput) DistillResult {
input.EpisodeID, overall, score.IS, score.SU, score.PA, score.VD, score.RU)
if overall < 0.7 && score.VD < 0.8 {
log.Printf("[distill] score below threshold for %s, skip", input.EpisodeID)
// 即使阈值未通过,仍然用 LLM 实体提取(如果有的话)
if len(llmResp.Entities) == 0 {
return DistillResult{}
// 人侧短事实(画像/心迹/牵挂/偏好/习惯/目标)是用户亲口陈述,
// LLM 打分天然偏低(短、无上下文),不应因分数被丢弃——否则"牧尘不在招商局工作"
// 这类关键事实永远到不了 memories,t6 永远比对不到。
if !isSoulfulShortFact(input.Content) {
log.Printf("[distill] score below threshold for %s, skip", input.EpisodeID)
// 即使阈值未通过,仍然用 LLM 实体提取(如果有的话)
if len(llmResp.Entities) == 0 {
return DistillResult{}
}
}
}
@@ -492,7 +497,9 @@ func (e *Engine) extractFacts(content string) ([]string, []Entity) {
seen[cleanChinese] = true
isTech := containsAny(cleanChinese, []string{"端口", "配置", "系统", "内存", "显卡", "服务", "记忆", "织忆", "飞书", "版本", "模型", "工具", "项目", "安装", "部署", "目录", "路径", "开发", "语言"})
etype := "entity"
if isTech { etype = "fact" }
if isTech {
etype = "fact"
}
entities = append(entities, Entity{
Name: cleanChinese, Type: etype, Properties: []string{"extracted"},
})
@@ -531,8 +538,12 @@ func isTechToken(s string) bool {
hasDigit := false
hasLetter := false
for _, r := range s {
if r >= '0' && r <= '9' { hasDigit = true }
if (r >= 'a' && r <= 'z') || (r >= 'A' && r <= 'Z') { hasLetter = true }
if r >= '0' && r <= '9' {
hasDigit = true
}
if (r >= 'a' && r <= 'z') || (r >= 'A' && r <= 'Z') {
hasLetter = true
}
}
return hasDigit && hasLetter
}
@@ -695,11 +706,11 @@ func (e *Engine) GetStatus() map[string]interface{} {
e.mu.Lock()
defer e.mu.Unlock()
return map[string]interface{}{
"queue_len": len(e.queue),
"batch_size": e.batchSize,
"last_distill": e.lastDistill.Format(time.RFC3339),
"daily_used": e.dailyUsed,
"daily_limit": e.dailyLimit,
"queue_len": len(e.queue),
"batch_size": e.batchSize,
"last_distill": e.lastDistill.Format(time.RFC3339),
"daily_used": e.dailyUsed,
"daily_limit": e.dailyLimit,
"daily_remaining": e.dailyLimit - e.dailyUsed,
}
}