@@ -538,6 +538,63 @@ test("GET /api/dsh-mneme/semantic reports configured/coverage/degraded", async (
538538 assert . equal ( noIndex . degraded , false ) ;
539539} ) ;
540540
541+ // --- distill-tx-embed-gap 后续:/semantic 区分「有向量」与「可检索」 -----------
542+
543+ test ( "GET /api/dsh-mneme/semantic reports per-dimension/per-model vector breakdown" , async ( ) => {
544+ const { routes, store, service } = setup ( undefined ) ;
545+ // 三种状态混进一个库:3 维 2 条、旧 4 维 1 条、不可解析 1 条;
546+ // 模型出身 m1#aaa 一行、m2#bbb 一行、无出身两行。
547+ const a = service . saveWithDedupe ( { type : "preference" , title : "同维A" , content : "x" } ) . memory ;
548+ const b = service . saveWithDedupe ( { type : "preference" , title : "同维B" , content : "x" } ) . memory ;
549+ const c = service . saveWithDedupe ( { type : "preference" , title : "旧维C" , content : "x" } ) . memory ;
550+ const d = service . saveWithDedupe ( { type : "preference" , title : "坏D" , content : "x" } ) . memory ;
551+ store . db . prepare ( "UPDATE memories SET embedding = ? WHERE id = ?" ) . run ( JSON . stringify ( [ 1 , 2 , 3 ] ) , a . id ) ;
552+ store . db . prepare ( "UPDATE memories SET embedding = ?, embedding_model = ? WHERE id = ?" ) . run ( JSON . stringify ( [ 1 , 2 , 3 ] ) , "m1#aaa" , b . id ) ;
553+ store . db . prepare ( "UPDATE memories SET embedding = ?, embedding_model = ? WHERE id = ?" ) . run ( JSON . stringify ( [ 1 , 2 , 3 , 4 ] ) , "m2#bbb" , c . id ) ;
554+ store . db . prepare ( "UPDATE memories SET embedding = ? WHERE id = ?" ) . run ( "not-json" , d . id ) ;
555+
556+ const sem = routes . find ( ( r ) => r . path === "/api/dsh-mneme/semantic" ) ;
557+ const res = new FakeRes ( ) ;
558+ await sem . handler ( req ( "/api/dsh-mneme/semantic" ) , res ) ;
559+ assert . equal ( res . statusCode , 200 ) ;
560+ const data = JSON . parse ( res . body ) ;
561+
562+ // 无 embedder → 无当前维度可比对:unsearchable 报 0、retrievable 报 null。
563+ assert . equal ( data . vectors . unsearchable , 0 ) ;
564+ assert . equal ( data . vectors . retrievable , null ) ;
565+ const dims = Object . fromEntries ( data . vectors . byDimension . map ( ( e ) => [ e . dimension , e . count ] ) ) ;
566+ assert . equal ( dims [ "3" ] , 2 ) ;
567+ assert . equal ( dims [ "4" ] , 1 ) ;
568+ assert . equal ( dims [ null ] , 1 , "unparseable vector counted under null" ) ;
569+ const models = Object . fromEntries ( data . vectors . byModel . map ( ( e ) => [ e . model , e . count ] ) ) ;
570+ assert . equal ( models [ "m1#aaa" ] , 1 ) ;
571+ assert . equal ( models [ "m2#bbb" ] , 1 ) ;
572+ assert . equal ( models [ null ] , 2 , "rows without provenance counted under null" ) ;
573+ } ) ;
574+
575+ test ( "GET /api/dsh-mneme/semantic counts stale-dimension rows as unsearchable with a matching embedder" , async ( ) => {
576+ // 核心回归锁:库里掺一条错维向量,端点必须报出 unsearchable ≥ 1。
577+ // 旧实现无论怎么混都只给 coverage,状态卡说不出「有向量但检索不到」。
578+ const { routes, store, service } = setup ( { name : "OpenAI" , embed ( ) { } , dimension : 3 } ) ;
579+ const a = service . saveWithDedupe ( { type : "preference" , title : "同维" , content : "x" } ) . memory ;
580+ const b = service . saveWithDedupe ( { type : "preference" , title : "错维" , content : "x" } ) . memory ;
581+ store . db . prepare ( "UPDATE memories SET embedding = ? WHERE id = ?" ) . run ( JSON . stringify ( [ 1 , 2 , 3 ] ) , a . id ) ;
582+ store . db . prepare ( "UPDATE memories SET embedding = ? WHERE id = ?" ) . run ( JSON . stringify ( [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 ] ) , b . id ) ;
583+ const sem = routes . find ( ( r ) => r . path === "/api/dsh-mneme/semantic" ) ;
584+ const res = new FakeRes ( ) ;
585+ await sem . handler ( req ( "/api/dsh-mneme/semantic" ) , res ) ;
586+ assert . equal ( res . statusCode , 200 ) ;
587+ const data = JSON . parse ( res . body ) ;
588+ assert . equal ( data . vectors . unsearchable , 1 , "the 8-dim row is not searchable against a 3-dim embedder" ) ;
589+ assert . equal ( data . vectors . retrievable , 1 , "the 3-dim row is searchable" ) ;
590+ // coverage 走 semantic.vectorIndex.getStats(本测试未挂 → 0),所以「有向量的行」
591+ // 用 byDimension 总和对照:两条都在「有向量」里,但只有一条可检索 —— 这正是
592+ // unsearchable 必须存在的原因。
593+ const embeddedRows = data . vectors . byDimension . reduce ( ( s , e ) => s + e . count , 0 ) ;
594+ assert . equal ( embeddedRows , 2 , "both rows have vectors" ) ;
595+ assert . equal ( data . vectors . unsearchable , 1 , "yet only one is retrievable" ) ;
596+ } ) ;
597+
541598// --- Bug8: llm-audit API (pagination + stats) --------------------------------
542599
543600test ( "GET /api/dsh-mneme/semantic/llm-audit returns paginated rows" , async ( ) => {
0 commit comments