fix: pass requested vector dimension to embedding API (Ollama + OpenAI)

The qwen3-embedding:4b model defaults to 2560-d vectors. Both
Ollama (client.embed()) and OpenAI support a dimensions parameter
to request a specific output size. This change threads the configured
qdrant.vectorSize through the AI provider layer so the model returns
vectors matching the Qdrant collection dimensions.

- AiApi.embeddings() now accepts optional dimensions parameter
- Ollama provider: switched from client.embeddings() to client.embed()
- OpenAI provider: passes dimensions to embeddings.create()
- VectorStoreService.getEmbedding() passes env.qdrant.vectorSize
- Added unit tests for dimension mismatch detection, collection
  creation, and search guards

gadget-code/src/services/vector-store.ts

@@ -252,6 +252,7 @@ class VectorStoreService extends DtpService {
     const response = await this.aiApi.embeddings(
       env.qdrant.embeddingModel,
       text,
+      env.qdrant.vectorSize,
     );
     return response.embedding;
   }

gadget-code/tests/vector-store.test.ts

@@ -0,0 +1,320 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+
+const mockEmbeddings = vi.hoisted(() => vi.fn());
+const mockFindById = vi.hoisted(() =>
+  vi.fn().mockResolvedValue({
+    _id: "test-provider-id",
+    name: "Test Provider",
+    apiType: "ollama",
+    baseUrl: "http://test:11434",
+    apiKey: "",
+  }),
+);
+
+// Create mock methods shared between QdrantClient mock and test code
+const mockQdrantMethods = vi.hoisted(() => ({
+  getCollections: vi.fn(),
+  createCollection: vi.fn(),
+  getCollection: vi.fn(),
+  search: vi.fn(),
+  upsert: vi.fn(),
+  delete: vi.fn(),
+}));
+
+vi.mock("../src/config/env.js", () => ({
+  default: {
+    NODE_ENV: "test",
+    INSTALL_DIR: "/tmp",
+    timezone: "UTC",
+    pkg: { name: "test", version: "0.0.0" },
+    site: {},
+    ai: { ollama: { apiUrl: "http://test:11434", apiKey: "" } },
+    auth: { jwtSecret: "test-secret" },
+    session: {
+      secret: "test-secret",
+      trustProxy: false,
+      cookie: { secure: false, sameSite: false },
+    },
+    google: { cse: { apiKey: "", engineId: "" } },
+    mongodb: { host: "localhost:27017", database: "test" },
+    qdrant: {
+      host: "localhost",
+      port: 6333,
+      collection: "test-collection",
+      providerId: "test-provider-id",
+      embeddingModel: "test-model",
+      vectorSize: 1024,
+    },
+    redis: {
+      host: "localhost",
+      port: 6379,
+      password: "",
+      keyPrefix: "test:",
+      lazyConnect: true,
+    },
+    minio: {
+      endpoint: "localhost",
+      port: 9000,
+      useSsl: false,
+      accessKey: "",
+      secretKey: "",
+      buckets: { uploads: "", images: "", videos: "", audios: "" },
+    },
+    user: { passwordSalt: "test-salt" },
+    https: { enabled: false, address: "localhost", port: 3443 },
+    socket: { maxHttpBufferSize: 1048576 },
+    frontend: { port: 5173 },
+    email: { enabled: false, smtp: {}, contact: {} },
+    log: {
+      https: { enabled: false },
+      console: { enabled: false },
+      file: { enabled: false },
+    },
+  },
+}));
+
+vi.mock("../src/models/ai-provider.js", () => ({
+  default: { findById: mockFindById },
+}));
+
+vi.mock("@gadget/ai", () => ({
+  createAiApi: vi.fn().mockReturnValue({
+    embeddings: mockEmbeddings,
+  }),
+}));
+
+vi.mock("@langchain/textsplitters", () => ({
+  RecursiveCharacterTextSplitter: vi.fn(function () {
+    return {
+      splitText: vi.fn().mockResolvedValue(["chunk1"]),
+    };
+  }),
+}));
+
+// QdrantClient mock constructor that returns the shared mock methods
+vi.mock("@qdrant/js-client-rest", () => ({
+  QdrantClient: vi.fn(function () {
+    return mockQdrantMethods;
+  }),
+}));
+
+import VectorStoreService from "../src/services/vector-store.js";
+
+const svc = VectorStoreService as unknown as {
+  _initialized: boolean;
+  _dimensionMismatch: boolean;
+  client: typeof mockQdrantMethods;
+  aiApi: {
+    embeddings: ReturnType<typeof vi.fn>;
+  };
+};
+
+function setMockDefaults() {
+  mockQdrantMethods.getCollections.mockResolvedValue({ collections: [] });
+  mockQdrantMethods.createCollection.mockResolvedValue(undefined);
+  mockQdrantMethods.getCollection.mockResolvedValue({});
+  mockQdrantMethods.search.mockResolvedValue([]);
+  mockQdrantMethods.upsert.mockResolvedValue(undefined);
+  mockQdrantMethods.delete.mockResolvedValue(undefined);
+}
+
+function ensureInstanceSetup() {
+  // For tests that don't call start(), manually assign client/aiApi
+  // since start() normally does this via new QdrantClient() + createAiApi()
+  if (!svc.client) {
+    (svc as Record<string, unknown>).client = mockQdrantMethods;
+  }
+  if (!svc.aiApi) {
+    (svc as Record<string, unknown>).aiApi = { embeddings: mockEmbeddings };
+  }
+}
+
+describe("VectorStoreService", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    setMockDefaults();
+
+    // Reset internal state on the instance
+    svc._initialized = false;
+    svc._dimensionMismatch = false;
+  });
+
+  describe("search", () => {
+    beforeEach(() => {
+      ensureInstanceSetup();
+    });
+
+    it("throws when service is not initialized", async () => {
+      svc._initialized = false;
+      svc._dimensionMismatch = false;
+
+      await expect(VectorStoreService.search("test query")).rejects.toThrow(
+        "VectorStoreService is not initialized",
+      );
+    });
+
+    it("throws when dimension mismatch flag is set", async () => {
+      svc._initialized = true;
+      svc._dimensionMismatch = true;
+
+      await expect(VectorStoreService.search("test query")).rejects.toThrow(
+        "Vector dimension mismatch: the Qdrant collection dimensions do not match the configured vectorSize (1024)",
+      );
+    });
+
+    it("throws when query embedding dimensions mismatch config", async () => {
+      svc._initialized = true;
+      svc._dimensionMismatch = false;
+
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(512).fill(0.1),
+        model: "test-model",
+      });
+
+      await expect(VectorStoreService.search("test query")).rejects.toThrow(
+        "Embedding dimension mismatch: model produced 512 dimensions, but collection expects 1024",
+      );
+    });
+
+    it("passes vectorSize dimensions to the AI API when embedding", async () => {
+      svc._initialized = true;
+      svc._dimensionMismatch = false;
+
+      mockQdrantMethods.search.mockResolvedValue([]);
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(1024).fill(0.1),
+        model: "test-model",
+      });
+
+      await VectorStoreService.search("test query");
+
+      expect(mockEmbeddings).toHaveBeenCalledWith(
+        "test-model",
+        "test query",
+        1024,
+      );
+    });
+
+    it("returns hydrated search results", async () => {
+      svc._initialized = true;
+      svc._dimensionMismatch = false;
+
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(1024).fill(0.1),
+        model: "test-model",
+      });
+
+      mockQdrantMethods.search.mockResolvedValue([
+        {
+          id: "point-1",
+          score: 0.95,
+          payload: {
+            content: "some content",
+            userId: "user-1",
+            projectId: "proj-1",
+            sessionId: "session-1",
+            turnId: "turn-1",
+            role: "user",
+            createdAt: "2026-01-01T00:00:00.000Z",
+          },
+        },
+      ]);
+
+      const results = await VectorStoreService.search("test query", undefined, 5);
+      expect(results).toHaveLength(1);
+      expect(results[0]).toMatchObject({
+        id: "point-1",
+        content: "some content",
+        score: 0.95,
+        userId: "user-1",
+      });
+    });
+  });
+
+  describe("start / ensureCollection", () => {
+    it("creates collection when it does not exist", async () => {
+      mockQdrantMethods.getCollections.mockResolvedValue({
+        collections: [],
+      });
+      mockQdrantMethods.createCollection.mockResolvedValue(undefined);
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(1024).fill(0.1),
+        model: "test-model",
+      });
+
+      await VectorStoreService.start();
+
+      expect(mockQdrantMethods.createCollection).toHaveBeenCalledWith(
+        "test-collection",
+        { vectors: { size: 1024, distance: "Cosine" } },
+      );
+      expect(svc._initialized).toBe(true);
+      expect(svc._dimensionMismatch).toBe(false);
+    });
+
+    it("detects dimension mismatch when existing collection has wrong size", async () => {
+      mockQdrantMethods.getCollections.mockResolvedValue({
+        collections: [{ name: "test-collection" }],
+      });
+      mockQdrantMethods.getCollection.mockResolvedValue({
+        config: {
+          params: {
+            vectors: { size: 768, distance: "Cosine" },
+          },
+        },
+      });
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(1024).fill(0.1),
+        model: "test-model",
+      });
+
+      await VectorStoreService.start();
+
+      expect(svc._initialized).toBe(true);
+      expect(svc._dimensionMismatch).toBe(true);
+    });
+
+    it("sets dimension mismatch when embedding model produces wrong dimensions", async () => {
+      mockQdrantMethods.getCollections.mockResolvedValue({
+        collections: [],
+      });
+      mockQdrantMethods.createCollection.mockResolvedValue(undefined);
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(512).fill(0.1),
+        model: "test-model",
+      });
+
+      await VectorStoreService.start();
+
+      expect(svc._initialized).toBe(true);
+      expect(svc._dimensionMismatch).toBe(true);
+    });
+
+    it("starts cleanly when everything matches", async () => {
+      mockQdrantMethods.getCollections.mockResolvedValue({
+        collections: [],
+      });
+      mockQdrantMethods.createCollection.mockResolvedValue(undefined);
+      mockEmbeddings.mockResolvedValue({
+        embedding: new Array(1024).fill(0.1),
+        model: "test-model",
+      });
+
+      await VectorStoreService.start();
+
+      expect(svc._dimensionMismatch).toBe(false);
+      expect(svc._initialized).toBe(true);
+    });
+
+    it("skips start when no providerId is configured", async () => {
+      const env = await import("../src/config/env.js");
+      const originalProviderId = env.default.qdrant.providerId;
+      env.default.qdrant.providerId = "";
+
+      await VectorStoreService.start();
+      expect(svc._initialized).toBe(false);
+
+      env.default.qdrant.providerId = originalProviderId;
+    });
+  });
+});

packages/ai/src/api.ts

@@ -204,7 +204,7 @@ export abstract class AiApi {
   /**
    * Generate an embedding vector for the given text.
    */
-  abstract embeddings(modelId: string, text: string): Promise<IAiEmbeddingResponse>;
+  abstract embeddings(modelId: string, text: string, dimensions?: number): Promise<IAiEmbeddingResponse>;
 
   /**
    * Forcefully abort any in-progress API request.

packages/ai/src/ollama.ts

@@ -369,8 +369,8 @@ export class OllamaAiApi extends AiApi {
     return chatResponse;
   }
 
-  async embeddings(modelId: string, text: string): Promise<IAiEmbeddingResponse> {
+  async embeddings(modelId: string, text: string, dimensions?: number): Promise<IAiEmbeddingResponse> {
-    const response = await this.client.embeddings({ model: modelId, prompt: text });
+    const response = await this.client.embed({ model: modelId, input: text, dimensions });
-    return { embedding: response.embedding, model: modelId };
+    return { embedding: response.embeddings[0]!, model: modelId };
   }
 }

packages/ai/src/openai.ts

@@ -648,11 +648,12 @@ export class OpenAiApi extends AiApi {
     };
   }
 
-  async embeddings(modelId: string, text: string): Promise<IAiEmbeddingResponse> {
+  async embeddings(modelId: string, text: string, dimensions?: number): Promise<IAiEmbeddingResponse> {
     const response = await this.client.embeddings.create({
       model: modelId,
       input: text,
+      dimensions,
     });
-    return { embedding: response.data[0].embedding, model: modelId };
+    return { embedding: response.data[0]!.embedding, model: modelId };
   }
 }