Vector Stores

VectorStore interface and implementations—InMemory, PgVector, Qdrant, MongoDB—for semantic search in RadarOS.

Vector Stores

Vector stores hold document embeddings and support similarity search. RadarOS provides four implementations. Choose based on your scale, infrastructure, and persistence needs.

---

VectorStore Interface

interface VectorStore {
  initialize(): Promise<void>;
  upsert(collection: string, doc: VectorDocument): Promise<void>;
  upsertBatch(collection: string, docs: VectorDocument[]): Promise<void>;
  search(
    collection: string,
    query: number[] | string,
    options?: VectorSearchOptions
  ): Promise<VectorSearchResult[]>;
  get(collection: string, id: string): Promise<VectorDocument | null>;
  delete(collection: string, id: string): Promise<void>;
  dropCollection(collection: string): Promise<void>;
  close(): Promise<void>;
}

Types

Type	Shape
VectorDocument	{ id, content, embedding?, metadata? }
VectorSearchOptions	{ topK?, minScore?, filter? }
VectorSearchResult	{ id, content, score, metadata? }

---

Implementations Comparison

Store	Dependencies	Initialize	Persistence	Best For
InMemoryVectorStore	None	No	No	Dev, tests
PgVectorStore	pg, pgvector	Yes	Yes	PostgreSQL users
QdrantVectorStore	@qdrant/js-client-rest	No	Yes	Dedicated vector DB
MongoDBVectorStore	mongodb	Yes	Yes	MongoDB / Atlas

---

InMemoryVectorStore

No dependencies. Data lost on restart. Ideal for development and tests.

import { InMemoryVectorStore, OpenAIEmbedding } from "@radaros/core";

const embedder = new OpenAIEmbedding({ apiKey: process.env.OPENAI_API_KEY });
const store = new InMemoryVectorStore(embedder);

await store.initialize(); // no-op, but call for consistency
await store.upsert("docs", {
  id: "1",
  content: "RadarOS is a TypeScript agent framework.",
});
const results = await store.search("docs", "What is RadarOS?");

---

PgVectorStore

Uses PostgreSQL with the pgvector extension. Requires pg and pgvector.

npm

npm install pg pgvector

import { PgVectorStore, OpenAIEmbedding } from "@radaros/core";

const embedder = new OpenAIEmbedding();
const store = new PgVectorStore(
  {
    connectionString: process.env.DATABASE_URL!,
    dimensions: 1536, // optional, from embedder if omitted
  },
  embedder
);

await store.initialize(); // Creates vector extension + tables
await store.upsert("docs", { id: "1", content: "..." });
const results = await store.search("docs", "query", { topK: 5 });

connectionString · string · required

PostgreSQL connection string.

tableName · string

Not used directly—collections map to tables. Table names are sanitized from collection names.

dimensions · number

Embedding dimensions. Defaults to embedder’s dimensions or 1536.

---

QdrantVectorStore

Uses Qdrant—a dedicated vector database. Requires @qdrant/js-client-rest.

npm

npm install @qdrant/js-client-rest

import { QdrantVectorStore, OpenAIEmbedding } from "@radaros/core";

const embedder = new OpenAIEmbedding();
const store = new QdrantVectorStore(
  {
    url: "http://localhost:6333",
    apiKey: process.env.QDRANT_API_KEY,
    dimensions: 1536,
  },
  embedder
);

await store.initialize();
await store.upsert("docs", { id: "1", content: "..." });
const results = await store.search("docs", "query", { topK: 5, minScore: 0.7 });

url · string · default: "http://localhost:6333"

Qdrant server URL.

apiKey · string

API key for Qdrant Cloud.

collectionName · string

Collections are created on first upsert. Pass collection name to upsert/search.

dimensions · number

Embedding dimensions. Defaults to embedder or 1536.

---

MongoDBVectorStore

Uses MongoDB. Supports Atlas Vector Search (when index exists) or local in-memory similarity. Requires mongodb.

npm

npm install mongodb

import { MongoDBVectorStore, OpenAIEmbedding } from "@radaros/core";

const embedder = new OpenAIEmbedding();
const store = new MongoDBVectorStore(
  {
    uri: process.env.MONGO_URI!,
    dbName: "radaros_vectors",
    indexName: "vector_index", // Atlas Search index name
    dimensions: 1536,
  },
  embedder
);

await store.initialize();
await store.upsert("docs", { id: "1", content: "..." });
const results = await store.search("docs", "query", { topK: 5 });

uri · string · required

MongoDB connection URI.

dbName · string · default: "radaros_vectors"

Database name.

collectionName · string

Collections are created on first upsert.

indexName · string · default: "vector_index"

Atlas Search index name for $vectorSearch. Must be pre-created for Atlas.

dimensions · number

Embedding dimensions.

---

EmbeddingProvider

All vector stores accept an optional EmbeddingProvider to compute embeddings when documents don’t include them. See Embeddings for OpenAI and Google providers.

// With embedder—documents without embedding get one automatically
const store = new InMemoryVectorStore(embedder);
await store.upsert("docs", { id: "1", content: "Hello" }); // embedding computed

// Without embedder—documents must include embedding
await store.upsert("docs", {
  id: "1",
  content: "Hello",
  embedding: [0.1, -0.2, ...],
});