mirror of
https://gitcode.com/JianFeeeee/HomeAgent.git
synced 2026-10-03 15:53:56 +00:00
refactor(memory): 拆除描述式媒体索引，媒体成为一等块并按原生向量融合
背景：此前媒体是靠「生成的描述文本」将就进记忆的——写 marker 进正文、 再由正则反解成 media_refs 与图库里的 type=Media 实体。这条链路有三个 致命缺陷：描述由异步模型生成（未生成前媒体等于不存在）、语义检索实质上 只搜描述文字、图库里的「媒体节点」是描述文本的投影而不是媒体本身。 本提交把这条链路整体拆除，媒体改为按自己的原生向量参与记忆： 一、描述链彻底删除（无残留、无兼容分支） - media.Item 去掉 Description/DescribedBy 与对应列； - 删除 Store.Describe / Store.Search / Store.Pending； - 删除 Agent.mediaDescribeLoop / describePendingMedia 与配置项 core.memory.media.describe_on_ingest； - SDK 侧 MediaAttachment 去掉 Description（见 SDK 仓独立提交）。 二、marker 机制删除，媒体归属改为结构化块边 - 删除 mediaMarkerLine/parseMediaMarkers/mediaEntityName/mediaTriplesFromText/ extractMediaDigests/sentenceWithMediaMarkers/docMediaContext； - memory.Triple 新增 MediaDigests 结构化字段；句子文本保持原样， 不再被 marker 污染； - 块以 sentence --contains--> block / document --contains--> block 结构边 挂到承载节点（新增 documents 表与 document 节点种类）； - 模型未给原句时用「主谓宾。」拼一句自然语言作落点，不造 marker 文本。 三、旧数据迁移（幂等） - 新增 GraphDB.MigrateLegacyMediaEntities：把 type=Media 的旧实体按短 digest 还原成原生块、挂回原句子、删除旧实体与描述关系；Agent 启动时执行； - CleanupOrphanedSentences 同时看关系引用与块边，避免把只靠块存活的句子 连同块边一起删掉。 四、向量融合：媒体按图本身被召回 - 新增 vector.FuseVectors（逐维求和 + L2 归一化）； - Doc.DenseVec = 文本向量 ⊕ 文档块的媒体向量（同 fingerprint 才融合）， 新增 Doc.DenseFP，指纹变化触发重算； - ContextEvent.DenseVec 同理融合事件块；事件新增 DenseFP，Prune 只在 同一统一空间内比稠密余弦； - 跨模态视觉路只召回「仍被某层记忆块持有」的媒体，CAS 全库字节不再 直接充当记忆检索结果。 五、同时纳入本分支既有的嵌入基础改造（此前工作区未提交，缺它 HEAD 不可构建） - internal/tfidf 懒回退包、千问三段式多模态 ONNX 空间的 Go 侧 （qwen/embedder.go、image.go、model_input.go）、CLIP 移除、 sdk.NewStore 分词器签名与调用点、embed 侧车 systemd 单元。 验证：go build ./... 、go vet ./...（含 -tags medialive）均通过； 在 HEAD 的独立 worktree 上重放本次暂存集后 go test -short ./internal/... 全部通过（端口冲突类用例在隔离环境中亦通过）。未提交工作区中与本改造 无关的改动（HarmonyOS、waiter、devicebridge、plan.md 等）。
This commit is contained in:
JianFeeeee
@ -28,13 +28,13 @@ func cleanQQTemplate(text string) string {
|
||||
}
|
||||
|
||||
type cleanTestEvent struct {
|
||||
idx int
|
||||
source string
|
||||
input string
|
||||
response string
|
||||
rawText string
|
||||
idx int
|
||||
source string
|
||||
input string
|
||||
response string
|
||||
rawText string
|
||||
cleanedText string
|
||||
topic string
|
||||
topic string
|
||||
}
|
||||
|
||||
func TestCleanStressPrecision(t *testing.T) {
|
||||
@ -59,53 +59,53 @@ func TestCleanStressPrecision(t *testing.T) {
|
||||
}
|
||||
t.Logf("topics: %v, events: %d", usedTopics, len(events))
|
||||
|
||||
for _, qTopic := range usedTopics {
|
||||
query := queryForTopic(qTopic)
|
||||
qVec := e.Vectorize(query)
|
||||
for _, qTopic := range usedTopics {
|
||||
query := queryForTopic(qTopic)
|
||||
qVec := e.Vectorize(query)
|
||||
|
||||
type scored struct {
|
||||
idx int
|
||||
topic string
|
||||
text string
|
||||
score float64
|
||||
}
|
||||
all := make([]scored, len(events))
|
||||
for i, ev := range events {
|
||||
text := ev.rawText
|
||||
if cleanMode {
|
||||
text = ev.cleanedText
|
||||
type scored struct {
|
||||
idx int
|
||||
topic string
|
||||
text string
|
||||
score float64
|
||||
}
|
||||
all := make([]scored, len(events))
|
||||
for i, ev := range events {
|
||||
text := ev.rawText
|
||||
if cleanMode {
|
||||
text = ev.cleanedText
|
||||
}
|
||||
vec := e.Vectorize(text)
|
||||
all[i] = scored{idx: i, topic: ev.topic, text: text, score: cosineSim(qVec, vec)}
|
||||
}
|
||||
sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
|
||||
|
||||
topK := len(usedTopics) * 2
|
||||
if topK > len(all) {
|
||||
topK = len(all)
|
||||
}
|
||||
|
||||
intraHits := 0
|
||||
for _, s := range all[:topK] {
|
||||
if s.topic == qTopic {
|
||||
intraHits++
|
||||
}
|
||||
}
|
||||
expected := countTopicEvents(events, qTopic)
|
||||
if expected > topK {
|
||||
expected = topK
|
||||
}
|
||||
recall := float64(intraHits) / float64(expected)
|
||||
|
||||
if recall < 0.3 {
|
||||
t.Logf(" [LOW] query=%q topK=%d intra=%d/%d recall=%.2f", qTopic, topK, intraHits, expected, recall)
|
||||
for _, s := range all[:8] {
|
||||
t.Logf(" [%.4f] %s", s.score, trimLen(s.text, 60))
|
||||
}
|
||||
} else {
|
||||
t.Logf(" [OK] query=%q topK=%d intra=%d/%d recall=%.2f", qTopic, topK, intraHits, expected, recall)
|
||||
}
|
||||
}
|
||||
vec := e.Vectorize(text)
|
||||
all[i] = scored{idx: i, topic: ev.topic, text: text, score: cosineSim(qVec, vec)}
|
||||
}
|
||||
sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
|
||||
|
||||
topK := len(usedTopics) * 2
|
||||
if topK > len(all) {
|
||||
topK = len(all)
|
||||
}
|
||||
|
||||
intraHits := 0
|
||||
for _, s := range all[:topK] {
|
||||
if s.topic == qTopic {
|
||||
intraHits++
|
||||
}
|
||||
}
|
||||
expected := countTopicEvents(events, qTopic)
|
||||
if expected > topK {
|
||||
expected = topK
|
||||
}
|
||||
recall := float64(intraHits) / float64(expected)
|
||||
|
||||
if recall < 0.3 {
|
||||
t.Logf(" [LOW] query=%q topK=%d intra=%d/%d recall=%.2f", qTopic, topK, intraHits, expected, recall)
|
||||
for _, s := range all[:8] {
|
||||
t.Logf(" [%.4f] %s", s.score, trimLen(s.text, 60))
|
||||
}
|
||||
} else {
|
||||
t.Logf(" [OK] query=%q topK=%d intra=%d/%d recall=%.2f", qTopic, topK, intraHits, expected, recall)
|
||||
}
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
@ -252,11 +252,11 @@ func genStressEvents(n int) []cleanTestEvent {
|
||||
keywords []string
|
||||
sources []string
|
||||
}{
|
||||
"大学招生": {[]string{"河南医药大学", "录取分数线", "专业排名", "高考志愿", "招生简章"}, []string{"qq", "qq", "agent"}},
|
||||
"老大私聊": {[]string{"老大私聊消息", "回复老大", "任务安排", "汇报工作", "收到"}, []string{"qq", "agent", "agent"}},
|
||||
"前端开发": {[]string{"前端组件封装", "页面路由配置", "界面布局设计", "交互逻辑开发", "代码调试优化"}, []string{"cli", "cli", "agent"}},
|
||||
"大学招生": {[]string{"河南医药大学", "录取分数线", "专业排名", "高考志愿", "招生简章"}, []string{"qq", "qq", "agent"}},
|
||||
"老大私聊": {[]string{"老大私聊消息", "回复老大", "任务安排", "汇报工作", "收到"}, []string{"qq", "agent", "agent"}},
|
||||
"前端开发": {[]string{"前端组件封装", "页面路由配置", "界面布局设计", "交互逻辑开发", "代码调试优化"}, []string{"cli", "cli", "agent"}},
|
||||
"服务器运维": {[]string{"反向代理配置", "容器部署方案", "证书续期", "数据库备份恢复", "监控告警处理"}, []string{"cli", "agent", "agent"}},
|
||||
"股票基金": {[]string{"基金定投策略", "股票涨跌分析", "理财收益计算", "市场行情分析", "投资风险管理"}, []string{"qq", "qq", "agent"}},
|
||||
"股票基金": {[]string{"基金定投策略", "股票涨跌分析", "理财收益计算", "市场行情分析", "投资风险管理"}, []string{"qq", "qq", "agent"}},
|
||||
}
|
||||
|
||||
for i := 0; i < n; i++ {
|
||||
@ -292,13 +292,13 @@ func genStressEvents(n int) []cleanTestEvent {
|
||||
cleaned := cleanEventText(src, input, response)
|
||||
raw := rawEventText(src, input, response)
|
||||
events = append(events, cleanTestEvent{
|
||||
idx: i,
|
||||
source: src,
|
||||
input: input,
|
||||
response: response,
|
||||
rawText: raw,
|
||||
idx: i,
|
||||
source: src,
|
||||
input: input,
|
||||
response: response,
|
||||
rawText: raw,
|
||||
cleanedText: cleaned,
|
||||
topic: tp,
|
||||
topic: tp,
|
||||
})
|
||||
}
|
||||
return events
|
||||
|
||||
Reference in New Issue
Block a user