diff --git a/src/__tests__/embeddings.test.ts b/src/__tests__/embeddings.test.ts
index 67b8005..871c9d1 100644
--- a/src/__tests__/embeddings.test.ts
+++ b/src/__tests__/embeddings.test.ts
@@ -152,3 +152,26 @@ describe("embeddingConfigHash", () => {
     );
   });
 });
+
+describe("prepareEmbeddingText task prefixes", () => {
+  const item = { title: "Fix login bug", body: "resolves the redirect loop", type: "pr" };
+
+  it("prefixes the clustering task for a model trained to expect one", () => {
+    const result = prepareEmbeddingText(item, "embeddinggemma");
+
+    expect(result.startsWith("task: clustering | query: ")).toBe(true);
+    expect(result).toContain("Fix login bug");
+  });
+
+  it("leaves output byte-identical for a model with no task prompt", () => {
+    expect(prepareEmbeddingText(item, "nomic-embed-text-v2-moe")).toBe(prepareEmbeddingText(item));
+  });
+
+  it("adds no prefix for an unrecognised model, so a new model cannot be silently mangled", () => {
+    expect(prepareEmbeddingText(item, "some-future-model")).toBe(prepareEmbeddingText(item));
+  });
+
+  it("matches a tagged model variant, since ollama slugs carry a :tag", () => {
+    expect(prepareEmbeddingText(item, "embeddinggemma:300m").startsWith("task: clustering | query: ")).toBe(true);
+  });
+});
diff --git a/src/benchmark.ts b/src/benchmark.ts
index 78950db..bc54d38 100644
--- a/src/benchmark.ts
+++ b/src/benchmark.ts
@@ -261,7 +261,7 @@ export async function runBenchmarkForModel(
 
     for (let i = 0; i < allItems.length; i += batchSize) {
       const batch = allItems.slice(i, i + batchSize);
-      const texts = batch.map((item) => prepareEmbeddingText(item));
+      const texts = batch.map((item) => prepareEmbeddingText(item, model));
 
       let embeddings: number[][];
       try {
diff --git a/src/embeddings.ts b/src/embeddings.ts
index 2716ab5..60637e0 100644
--- a/src/embeddings.ts
+++ b/src/embeddings.ts
@@ -396,7 +396,35 @@ export async function createEmbeddingProvider(config: ProviderConfig): Promise<E
 // the embedding config hash so a format change invalidates cached embeddings and
 // the scan warns to re-embed, instead of silently mixing old and new text vectors.
 // v1: "Pull Request:"/"Issue:" type prefix. v2: no prefix.
-export const EMBEDDING_TEXT_VERSION = 2;
+// v3: per-model task prefix (see TASK_PREFIXES). Models with no entry emit
+// byte-identical text to v2, but the version still bumps: the model name alone
+// cannot distinguish vectors cached before a prefix existed for that model.
+export const EMBEDDING_TEXT_VERSION = 3;
+
+/**
+ * Task instructions some models are trained to receive. Omitting the prompt on
+ * a model that expects one measurably degrades it, so benchmarking such a model
+ * without this would under-report it and wrongly favour the incumbent.
+ *
+ * Keyed by ollama slug with the `:tag` stripped. Deliberately an allowlist: an
+ * unknown model gets no prefix, because inventing an instruction format a model
+ * was not trained on is worse than sending none.
+ *
+ * Clustering rather than retrieval: this pipeline compares items to each other
+ * to find duplicates. It has no query/document asymmetry.
+ */
+const TASK_PREFIXES: Readonly<Record<string, string>> = Object.freeze({
+  // https://ai.google.dev/gemma/docs/embeddinggemma — prompt set includes
+  // "task: clustering | query: " for grouping semantically similar text.
+  embeddinggemma: "task: clustering | query: ",
+});
+
+/** Ollama slugs carry a `:tag` (embeddinggemma:300m); the prompt is a property
+ * of the model family, not the quantisation. */
+export function taskPrefixFor(model?: string): string {
+  if (!model) return "";
+  return TASK_PREFIXES[model.split(":")[0].trim().toLowerCase()] ?? "";
+}
 
 export type EmbeddingVectorGeneration = "native" | "provider-selected-v1" | "local-truncation-v1";
 
@@ -441,11 +469,13 @@ export function effectiveEmbeddingConfigHash(
   );
 }
 
-export function prepareEmbeddingText(item: { title: string; body: string; type: string }): string {
+export function prepareEmbeddingText(item: { title: string; body: string; type: string }, model?: string): string {
   // No type prefix. A leading "Pull Request:" / "Issue:" token systematically
   // pushes an issue away from its own fix PR in embedding space, which fragments
   // a single bug across separate clusters. Only takes effect after a re-embed.
   const title = (item.title || "Untitled").trim();
   const body = (item.body || "").trim().slice(0, 2000);
-  return body ? `${title}\n\n${body}` : title;
+  const text = body ? `${title}\n\n${body}` : title;
+  // Model-specific task instruction, empty for models that do not use one.
+  return `${taskPrefixFor(model)}${text}`;
 }
diff --git a/src/pipeline.ts b/src/pipeline.ts
index 206f0cb..f008128 100644
--- a/src/pipeline.ts
+++ b/src/pipeline.ts
@@ -84,11 +84,14 @@ export async function reEmbedStoredItems(
   for (let i = 0; i < items.length; i += batchSize) {
     const batch = items.slice(i, i + batchSize);
     const texts = batch.map((item) =>
-      prepareEmbeddingText({
-        title: item.title,
-        body: item.bodySnippet,
-        type: item.type,
-      }),
+      prepareEmbeddingText(
+        {
+          title: item.title,
+          body: item.bodySnippet,
+          type: item.type,
+        },
+        providerConfig.model,
+      ),
     );
     const embeddings = await embedder.embedBatch(texts);
     for (let j = 0; j < batch.length; j++) {
@@ -263,7 +266,7 @@ export async function runScan(
 
   for (let i = 0; i < newItems.length; i += BATCH_SIZE) {
     const batch = newItems.slice(i, i + BATCH_SIZE);
-    const texts = batch.map((item) => prepareEmbeddingText(item));
+    const texts = batch.map((item) => prepareEmbeddingText(item, env.EMBEDDING_MODEL));
     const embedWithRetry = async (input: string[]): Promise<number[][]> => {
       for (let attempt = 0; attempt < 3; attempt++) {
         try {