diff --git a/src/selfhost/vectorize.ts b/src/selfhost/vectorize.ts
index dafb7b1ed7..30bc0ec501 100644
--- a/src/selfhost/vectorize.ts
+++ b/src/selfhost/vectorize.ts
@@ -61,12 +61,34 @@ export function createSqliteVectorize(driver: SqliteDriver): Vectorize {
       const { rows } = opts.namespace
         ? driver.query(`SELECT id, embedding, metadata FROM ${TABLE} WHERE namespace=?`, [opts.namespace])
         : driver.query(`SELECT id, embedding, metadata FROM ${TABLE}`, []);
-      const scored: Match[] = rows.map((r) => {
+      // #8766: a stored vector whose width differs from the query's (an operator swapped embedding models
+      // without a full reindex) is SKIPPED, never scored — cosineSimilarity's Math.min would otherwise
+      // silently compute a wrong similarity over the shorter prefix. Skipping degrades to fewer candidates,
+      // the same fail-safe posture the RAG pipeline already takes everywhere else; Qdrant hard-fails this at
+      // the collection level and pgvector raises, so this brings the SQLite adapter to parity. One WARN per
+      // query (not per row) so a mixed-width store is visible without flooding the log.
+      const mismatchedWidths = new Set<number>();
+      const scored: Match[] = [];
+      for (const r of rows) {
         const values = JSON.parse(r.embedding as string) as number[];
+        if (values.length !== vector.length) {
+          mismatchedWidths.add(values.length);
+          continue;
+        }
         const metadata = r.metadata ? (JSON.parse(r.metadata as string) as Record<string, unknown>) : undefined;
         const score = cosineSimilarity(vector, values);
-        return metadata === undefined ? { id: r.id as string, score } : { id: r.id as string, score, metadata };
-      });
+        scored.push(metadata === undefined ? { id: r.id as string, score } : { id: r.id as string, score, metadata });
+      }
+      if (mismatchedWidths.size > 0) {
+        console.warn(
+          JSON.stringify({
+            event: "sqlite_vectorize_dimension_mismatch",
+            queryWidth: vector.length,
+            storedWidths: [...mismatchedWidths].sort((a, b) => a - b),
+            detail: "mismatched-width rows skipped — reindex after an embedding-model change",
+          }),
+        );
+      }
       scored.sort((a, b) => b.score - a.score);
       return { matches: scored.slice(0, opts.topK ?? 12) };
     },
diff --git a/test/unit/selfhost-vectorize.test.ts b/test/unit/selfhost-vectorize.test.ts
index 742b6c4f08..ed2a9c82a4 100644
--- a/test/unit/selfhost-vectorize.test.ts
+++ b/test/unit/selfhost-vectorize.test.ts
@@ -1,5 +1,5 @@
 import { DatabaseSync } from "node:sqlite";
-import { describe, expect, it } from "vitest";
+import { afterEach, describe, expect, it, vi } from "vitest";
 import { nodeSqliteDriver } from "../../src/selfhost/d1-adapter";
 import { cosineSimilarity, createSqliteVectorize } from "../../src/selfhost/vectorize";
 
@@ -83,3 +83,50 @@ describe("createSqliteVectorize (#979 local RAG)", () => {
     expect(res.matches.map((m) => m.id)).toContain("n2");
   });
 });
+
+describe("dimension-mismatch guard (#8766)", () => {
+  afterEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  it("SKIPS a stored vector whose width differs from the query's — never scores over a truncated prefix", async () => {
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => undefined);
+    const v = makeVectorize();
+    await v.upsert([
+      { id: "ok", values: [1, 0], namespace: "n" },
+      { id: "wide", values: [1, 0, 0, 0], namespace: "n" }, // stale row from a different-dimension model
+    ]);
+    const { matches } = await v.query([1, 0], { topK: 10, namespace: "n" });
+    expect(matches.map((m) => m.id)).toEqual(["ok"]);
+    // One WARN per query, naming both widths.
+    expect(warn).toHaveBeenCalledTimes(1);
+    const logged = JSON.parse(warn.mock.calls[0]![0] as string);
+    expect(logged).toMatchObject({ event: "sqlite_vectorize_dimension_mismatch", queryWidth: 2, storedWidths: [4] });
+  });
+
+  it("logs NOTHING when every stored width matches (the steady state stays silent)", async () => {
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => undefined);
+    const v = makeVectorize();
+    await v.upsert([{ id: "a", values: [1, 0], namespace: "n" }]);
+    const { matches } = await v.query([1, 0], { topK: 10, namespace: "n" });
+    expect(matches).toHaveLength(1);
+    expect(warn).not.toHaveBeenCalled();
+  });
+
+  it("dedupes the warned widths and still returns matching rows sorted by score", async () => {
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => undefined);
+    const v = makeVectorize();
+    await v.upsert([
+      { id: "best", values: [1, 0], namespace: "n" },
+      { id: "worse", values: [0.5, 0.5], namespace: "n" },
+      { id: "w4a", values: [1, 0, 0, 0], namespace: "n" },
+      { id: "w4b", values: [0, 1, 0, 0], namespace: "n" },
+      { id: "w3", values: [1, 0, 0], namespace: "n" },
+    ]);
+    const { matches } = await v.query([1, 0], { topK: 10, namespace: "n" });
+    expect(matches.map((m) => m.id)).toEqual(["best", "worse"]);
+    const logged = JSON.parse((vi.mocked(console.warn).mock.calls[0]!)[0] as string);
+    expect(logged.storedWidths).toEqual([3, 4]);
+    expect(warn).toHaveBeenCalledTimes(1);
+  });
+});