Plan değişimi P438

(@incTest bit,@types1 nvarchar(4000),@sid int,@dbName nvarchar(4000),@tag nvarchar(4000),@model nvarchar(4000),@q vector(1024),@top int)SELECT TOP (@top) c.ChunkId, c.DocId, d.SourceType, d.SourceRef, d.Title, c.Heading, c.Text,
       VECTOR_DISTANCE('cosine', c.Embedding, @q) AS Score
FROM rag.Chunk AS c
JOIN rag.Document AS d ON d.DocId = c.DocId
WHERE (d.IsActive = 1 OR (@incTest = 1 AND d.ExclusionReason = 'TestDb')) AND d.SourceType IN (@types1)
AND (@sid IS NULL OR d.ServerID = @sid) AND (@dbName IS NULL OR d.DatabaseName = @dbName)
AND (@tag IS NULL OR EXISTS (SELECT 1 FROM STRING_SPLIT(d.Tags, ',') AS t WHERE LTRIM(RTRIM(t.value)) = @tag)) AND c.Embedding IS NOT NULL AND c.EmbeddingModel = @model
ORDER BY Score, c.ChunkId

Plan change case #P438 | DB: DB-11 | Improvement | Confidence: Low (review pending)
Query: (@incTest bit,@types1 nvarchar(?),@sid int,@dbName nvarchar(?),@tag nvarchar(?),@model nvarchar(?),@q vector(?),@top int)SELECT TOP (@top) c.ChunkId, c.DocId, d.SourceType, d.SourceRef, d.Title, c.Heading, c.Text, VECTOR_DISTANCE(?, c.Embedding, @q) AS Score FROM rag.Chunk AS c JOIN rag.Document AS d ON d.DocId = c.DocId WHERE (d.IsActive = ? OR (@incTest = ? AND d.ExclusionReason = ?)) AND d.SourceType IN (@types1) AND (@sid IS NULL OR d.ServerID = @sid) AND (@dbName IS NULL OR d.DatabaseName = @dbName) AND (@tag IS NULL OR EXISTS (SELECT ? FROM STRING_SPLIT(d.Tags, ?) AS t WHERE LTRIM(RTRIM(t.value)) = @tag)) AND c.Embedding IS NOT NULL AND c.EmbeddingModel = @model ORDER BY Score, c.ChunkId
Change: 2026-10-07 05:00 UTC, plan 64747 → 64778, pattern: Transition
Impact (Query Store, per execution): duration 33.9 → 24.3 ms (-28%); CPU 33.8 → 24.3 ms (-28%); reads 14,589 → 10,653 (-27%); rows 18.44 → 17.65 (-4%); aborted 1 → 0
Plan difference: no operator-level difference
Candidate cause: STATS — Statistic rag.Chunk.PK_Chunk was updated 2026-10-06 22:04 (sampled 42%); 4 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Chunk.UQ_Chunk_DocSeq was updated 2026-10-06 22:04 (sampled 52.7%); 4 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Chunk._WA_Sys_00000007_49309088 was updated 2026-10-06 22:05 (sampled 52.7%); 3 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Chunk._WA_Sys_00000003_49309088 was updated 2026-10-06 22:04 (sampled 42%); 4 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Document.PK_Document was updated 2026-10-06 01:16 (sampled 100%); the plans use this statistic.
Candidate cause: STATS — Statistic rag.Document.UQ_Document_Source was updated 2026-10-06 04:00 (sampled 100%); 2 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Document._WA_Sys_0000000A_428392F9 was updated 2026-10-06 19:19 (sampled 100%); 3 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Document._WA_Sys_00000005_428392F9 was updated 2026-10-06 22:02 (sampled 100%); 3 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Document._WA_Sys_0000000E_428392F9 was updated 2026-10-06 05:48 (sampled 100%); 2 updates in the window; the plans use this statistic.
Candidate cause: STATS — Statistic rag.Document._WA_Sys_00000007_428392F9 was updated 2026-10-06 22:07 (sampled 100%); the plans use this statistic.
Counter-evidence: HIGH_VARIANCE — The old plan's duration standard deviation is 104% of the average: the average is not reliable.
Counter-evidence: HIGH_VARIANCE — The new plan's duration standard deviation is 132% of the average: the average is not reliable.
Counter-evidence: TREND_BEFORE — The old plan was already drifting before the change (avg 34.9 → 7.8 ms): part of the duration difference is independent of the plan change.
Counter-evidence: SAME_SHAPE — The two plans have the same operator tree (operators, tables, indexes): the plan did not change; the difference is only in the compiled parameter values and row estimates. The slowness may come from the amount of data this parameter value processes or from waits, not from plan choice.
Confidence rationale: Low: strong counter-evidence (SAME_SHAPE).
Note: this is a candidate cause, not proof of the root cause.