memory: normalize Gemini embeddings (#43409)

Merged via squash.

Prepared head SHA: 70613e0225
Co-authored-by: gumadeiras <5599352+gumadeiras@users.noreply.github.com>
Co-authored-by: gumadeiras <5599352+gumadeiras@users.noreply.github.com>
Reviewed-by: @gumadeiras

CHANGELOG.md

@@ -12,7 +12,7 @@ Docs: https://docs.openclaw.ai
 - Exec/child commands: mark child command environments with `OPENCLAW_CLI` so subprocesses can detect when they were launched from the OpenClaw CLI. (#41411) Thanks @vincentkoc.
 - iOS/Home canvas: add a bundled welcome screen with a live agent overview that refreshes on connect, reconnect, and foreground return, and move the compact connection pill off the top-left canvas overlay. (#42456) Thanks @ngutman.
 - iOS/Home canvas: replace floating controls with a docked toolbar, make the bundled home scaffold adapt to smaller phones, and open chat in the resolved main session instead of a synthetic `ios` session. (#42456) Thanks @ngutman.
-- Memory/Gemini: add `gemini-embedding-2-preview` memory-search support with configurable output dimensions and automatic reindexing when the configured dimensions change. (#42501) thanks @BillChirico.
+- Memory/Gemini: add `gemini-embedding-2-preview` memory-search support with configurable output dimensions and automatic reindexing when the configured dimensions change. (#42501) Thanks @BillChirico and @gumadeiras.
 - Discord/auto threads: add `autoArchiveDuration` channel config for auto-created threads so Discord thread archiving can stay at 1 hour, 1 day, 3 days, or 1 week instead of always using the 1-hour default. (#35065) Thanks @davidguttman.
 - OpenCode/onboarding: add new OpenCode Go provider, treat Zen and Go as one OpenCode setup in the wizard/docs while keeping the runtime providers split, store one shared OpenCode key for both profiles, and stop overriding the built-in `opencode-go` catalog routing. (#42313) Thanks @ImLukeF and @vincentkoc.
 - macOS/chat UI: add a chat model picker, persist explicit thinking-level selections across relaunch, and harden provider-aware session model sync for the shared chat composer. (#42314) Thanks @ImLukeF.
@@ -181,6 +181,7 @@ Docs: https://docs.openclaw.ai
 - SecretRef/models: harden custom/provider secret persistence and reuse across models.json snapshots, merge behavior, runtime headers, and secret audits. (#42554) Thanks @joshavant.
 - macOS/browser proxy: serialize non-GET browser proxy request bodies through `AnyCodable.foundationValue` so nested JSON bodies no longer crash the macOS app with `Invalid type in JSON write (__SwiftValue)`. (#43069) Thanks @Effet.
 - CLI/skills tables: keep terminal table borders aligned for wide graphemes, use full reported terminal width, and switch a few ambiguous skill icons to Terminal-safe emoji so `openclaw skills` renders more consistently in Terminal.app and iTerm. Thanks @vincentkoc.
+- Memory/Gemini: normalize returned Gemini embeddings across direct query, direct batch, and async batch paths so memory search uses consistent vector handling for Gemini too. (#43409) Thanks @gumadeiras.
 
 ## 2026.3.7
 

src/memory/batch-gemini.test.ts

@@ -1,6 +1,10 @@
 import { afterEach, beforeAll, describe, expect, it, vi } from "vitest";
 import type { GeminiEmbeddingClient } from "./embeddings-gemini.js";
 
+function magnitude(values: number[]) {
+  return Math.sqrt(values.reduce((sum, value) => sum + value * value, 0));
+}
+
 describe("runGeminiEmbeddingBatches", () => {
   let runGeminiEmbeddingBatches: typeof import("./batch-gemini.js").runGeminiEmbeddingBatches;
 
@@ -56,7 +60,7 @@ describe("runGeminiEmbeddingBatches", () => {
         return new Response(
           JSON.stringify({
             key: "req-1",
-            response: { embedding: { values: [0.1, 0.2, 0.3] } },
+            response: { embedding: { values: [3, 4] } },
           }),
           {
             status: 200,
@@ -88,7 +92,11 @@ describe("runGeminiEmbeddingBatches", () => {
       concurrency: 1,
     });
 
-    expect(results.get("req-1")).toEqual([0.1, 0.2, 0.3]);
+    const embedding = results.get("req-1");
+    expect(embedding).toBeDefined();
+    expect(embedding?.[0]).toBeCloseTo(0.6, 5);
+    expect(embedding?.[1]).toBeCloseTo(0.8, 5);
+    expect(magnitude(embedding ?? [])).toBeCloseTo(1, 5);
     expect(fetchMock).toHaveBeenCalledTimes(3);
   });
 });

src/memory/batch-gemini.ts

@@ -4,6 +4,7 @@ import {
   type EmbeddingBatchExecutionParams,
 } from "./batch-runner.js";
 import { buildBatchHeaders, normalizeBatchBaseUrl } from "./batch-utils.js";
+import { sanitizeAndNormalizeEmbedding } from "./embedding-vectors.js";
 import { debugEmbeddingsLog } from "./embeddings-debug.js";
 import type { GeminiEmbeddingClient, GeminiTextEmbeddingRequest } from "./embeddings-gemini.js";
 import { hashText } from "./internal.js";
@@ -346,7 +347,9 @@ export async function runGeminiEmbeddingBatches(
           errors.push(`${customId}: ${line.response.error.message}`);
           continue;
         }
-        const embedding = line.embedding?.values ?? line.response?.embedding?.values ?? [];
+        const embedding = sanitizeAndNormalizeEmbedding(
+          line.embedding?.values ?? line.response?.embedding?.values ?? [],
+        );
         if (embedding.length === 0) {
           errors.push(`${customId}: empty embedding`);
           continue;

src/memory/embedding-vectors.ts

@@ -0,0 +1,8 @@
+export function sanitizeAndNormalizeEmbedding(vec: number[]): number[] {
+  const sanitized = vec.map((value) => (Number.isFinite(value) ? value : 0));
+  const magnitude = Math.sqrt(sanitized.reduce((sum, value) => sum + value * value, 0));
+  if (magnitude < 1e-10) {
+    return sanitized;
+  }
+  return sanitized.map((value) => value / magnitude);
+}

src/memory/embeddings-gemini.test.ts

@@ -44,6 +44,10 @@ function parseFetchBody(fetchMock: { mock: { calls: unknown[][] } }, callIndex =
   return JSON.parse((init?.body as string) ?? "{}") as Record<string, unknown>;
 }
 
+function magnitude(values: number[]) {
+  return Math.sqrt(values.reduce((sum, value) => sum + value * value, 0));
+}
+
 afterEach(() => {
   vi.resetAllMocks();
   vi.unstubAllGlobals();
@@ -224,6 +228,25 @@ describe("gemini-embedding-2-preview provider", () => {
     expect(body.content).toEqual({ parts: [{ text: "test query" }] });
   });
 
+  it("normalizes embedQuery response vectors", async () => {
+    const fetchMock = createGeminiFetchMock([3, 4]);
+    vi.stubGlobal("fetch", fetchMock);
+    mockResolvedProviderKey();
+
+    const { provider } = await createGeminiEmbeddingProvider({
+      config: {} as never,
+      provider: "gemini",
+      model: "gemini-embedding-2-preview",
+      fallback: "none",
+    });
+
+    const embedding = await provider.embedQuery("test query");
+
+    expect(embedding[0]).toBeCloseTo(0.6, 5);
+    expect(embedding[1]).toBeCloseTo(0.8, 5);
+    expect(magnitude(embedding)).toBeCloseTo(1, 5);
+  });
+
   it("includes outputDimensionality in embedBatch request", async () => {
     const fetchMock = createGeminiBatchFetchMock(2);
     vi.stubGlobal("fetch", fetchMock);
@@ -255,6 +278,28 @@ describe("gemini-embedding-2-preview provider", () => {
     ]);
   });
 
+  it("normalizes embedBatch response vectors", async () => {
+    const fetchMock = createGeminiBatchFetchMock(2, [3, 4]);
+    vi.stubGlobal("fetch", fetchMock);
+    mockResolvedProviderKey();
+
+    const { provider } = await createGeminiEmbeddingProvider({
+      config: {} as never,
+      provider: "gemini",
+      model: "gemini-embedding-2-preview",
+      fallback: "none",
+    });
+
+    const embeddings = await provider.embedBatch(["text1", "text2"]);
+
+    expect(embeddings).toHaveLength(2);
+    for (const embedding of embeddings) {
+      expect(embedding[0]).toBeCloseTo(0.6, 5);
+      expect(embedding[1]).toBeCloseTo(0.8, 5);
+      expect(magnitude(embedding)).toBeCloseTo(1, 5);
+    }
+  });
+
   it("respects custom outputDimensionality", async () => {
     const fetchMock = createGeminiFetchMock();
     vi.stubGlobal("fetch", fetchMock);
@@ -310,6 +355,28 @@ describe("gemini-embedding-2-preview provider", () => {
     ).rejects.toThrow(/Invalid outputDimensionality 512/);
   });
 
+  it("sanitizes non-finite values before normalization", async () => {
+    const fetchMock = createGeminiFetchMock([
+      1,
+      Number.NaN,
+      Number.POSITIVE_INFINITY,
+      Number.NEGATIVE_INFINITY,
+    ]);
+    vi.stubGlobal("fetch", fetchMock);
+    mockResolvedProviderKey();
+
+    const { provider } = await createGeminiEmbeddingProvider({
+      config: {} as never,
+      provider: "gemini",
+      model: "gemini-embedding-2-preview",
+      fallback: "none",
+    });
+
+    const embedding = await provider.embedQuery("test");
+
+    expect(embedding).toEqual([1, 0, 0, 0]);
+  });
+
   it("uses correct endpoint URL", async () => {
     const fetchMock = createGeminiFetchMock();
     vi.stubGlobal("fetch", fetchMock);

src/memory/embeddings-gemini.ts

@@ -5,6 +5,7 @@ import {
 import { requireApiKey, resolveApiKeyForProvider } from "../agents/model-auth.js";
 import { parseGeminiAuth } from "../infra/gemini-auth.js";
 import type { SsrFPolicy } from "../infra/net/ssrf.js";
+import { sanitizeAndNormalizeEmbedding } from "./embedding-vectors.js";
 import { debugEmbeddingsLog } from "./embeddings-debug.js";
 import type { EmbeddingProvider, EmbeddingProviderOptions } from "./embeddings.js";
 import { buildRemoteBaseUrlPolicy, withRemoteHttpResponse } from "./remote-http.js";
@@ -222,7 +223,7 @@ export async function createGeminiEmbeddingProvider(
       apiKeys: client.apiKeys,
       execute: (apiKey) => fetchWithGeminiAuth(apiKey, embedUrl, body),
     });
-    return payload.embedding?.values ?? [];
+    return sanitizeAndNormalizeEmbedding(payload.embedding?.values ?? []);
   };
 
   const embedBatch = async (texts: string[]): Promise<number[][]> => {
@@ -244,7 +245,7 @@ export async function createGeminiEmbeddingProvider(
       execute: (apiKey) => fetchWithGeminiAuth(apiKey, batchUrl, batchBody),
     });
     const embeddings = Array.isArray(payload.embeddings) ? payload.embeddings : [];
-    return texts.map((_, index) => embeddings[index]?.values ?? []);
+    return texts.map((_, index) => sanitizeAndNormalizeEmbedding(embeddings[index]?.values ?? []));
   };
 
   return {

src/memory/embeddings-ollama.ts

@@ -3,6 +3,7 @@ import { resolveOllamaApiBase } from "../agents/ollama-models.js";
 import { formatErrorMessage } from "../infra/errors.js";
 import type { SsrFPolicy } from "../infra/net/ssrf.js";
 import { normalizeOptionalSecretInput } from "../utils/normalize-secret-input.js";
+import { sanitizeAndNormalizeEmbedding } from "./embedding-vectors.js";
 import { normalizeEmbeddingModelWithPrefixes } from "./embeddings-model-normalize.js";
 import type { EmbeddingProvider, EmbeddingProviderOptions } from "./embeddings.js";
 import { buildRemoteBaseUrlPolicy, withRemoteHttpResponse } from "./remote-http.js";
@@ -19,15 +20,6 @@ type OllamaEmbeddingClientConfig = Omit<OllamaEmbeddingClient, "embedBatch">;
 
 export const DEFAULT_OLLAMA_EMBEDDING_MODEL = "nomic-embed-text";
 
-function sanitizeAndNormalizeEmbedding(vec: number[]): number[] {
-  const sanitized = vec.map((value) => (Number.isFinite(value) ? value : 0));
-  const magnitude = Math.sqrt(sanitized.reduce((sum, value) => sum + value * value, 0));
-  if (magnitude < 1e-10) {
-    return sanitized;
-  }
-  return sanitized.map((value) => value / magnitude);
-}
-
 function normalizeOllamaModel(model: string): string {
   return normalizeEmbeddingModelWithPrefixes({
     model,

src/memory/embeddings.ts

@@ -4,6 +4,7 @@ import type { OpenClawConfig } from "../config/config.js";
 import type { SecretInput } from "../config/types.secrets.js";
 import { formatErrorMessage } from "../infra/errors.js";
 import { resolveUserPath } from "../utils.js";
+import { sanitizeAndNormalizeEmbedding } from "./embedding-vectors.js";
 import {
   createGeminiEmbeddingProvider,
   type GeminiEmbeddingClient,
@@ -18,15 +19,6 @@ import { createOpenAiEmbeddingProvider, type OpenAiEmbeddingClient } from "./emb
 import { createVoyageEmbeddingProvider, type VoyageEmbeddingClient } from "./embeddings-voyage.js";
 import { importNodeLlamaCpp } from "./node-llama.js";
 
-function sanitizeAndNormalizeEmbedding(vec: number[]): number[] {
-  const sanitized = vec.map((value) => (Number.isFinite(value) ? value : 0));
-  const magnitude = Math.sqrt(sanitized.reduce((sum, value) => sum + value * value, 0));
-  if (magnitude < 1e-10) {
-    return sanitized;
-  }
-  return sanitized.map((value) => value / magnitude);
-}
-
 export type { GeminiEmbeddingClient } from "./embeddings-gemini.js";
 export type { MistralEmbeddingClient } from "./embeddings-mistral.js";
 export type { OpenAiEmbeddingClient } from "./embeddings-openai.js";