Official Resources

Documentation Changelog GA Announcement Bedrock Integration ENN Blog Quantization Blog Spring Data Blog

Key Features

  • Embedding Storage & Dimensions: Dense float32, int8, int1, and quantized vectors stored as regular document fields. Max 8192 dimensions as of June 2025.
  • Dual Index Types: ANN via tunable HNSW (efConstruction, efSearch, M) and ENN (exact) for ≤10k vectors or highly-filtered subsets.
  • Advanced Query API: $vectorSearch aggregation stage (MongoDB ≥6.0.11 for ANN, ≥7.0.10 for ENN) with pre-filter using standard MQL.
  • Vector Quantization: Scalar (int8) cuts RAM ~75%; Binary (int1) up to ~97% with automatic rescoring for binary vectors.
  • Search Nodes & Scalability: Search Nodes (GA) isolate vector workload, improve latency up to 60%. Horizontal scaling via sharding.
  • Security & Encryption: Field-Level Encryption & Queryable Encryption compatible with vector indexes for regulated environments.
  • Multi-Provider Support: Any provider ≤8192 dims (OpenAI, Cohere, Bedrock, Voyage, Jina, Nomic, etc.) supported.

Code Examples

Vector Search Index Creation

json 
{
  "fields": [
    {
      "type": "vector",
      "path": "embedding",
      "numDimensions": 1536,
      "similarity": "cosine"
    },
    {
      "type": "filter",
      "path": "tenantId"
    }
  ]
}

ANN with Pre-Filter Query

javascript 
db.movies.aggregate([
  {
    $vectorSearch: {
      index: "vector_index",
      path: "embedding",
      queryVector: [0.12, -0.34, ...], // 1536-dim
      numCandidates: 150,
      limit: 10,
      filter: { tenantId: { $eq: "acme" } }
    }
  }
])

Exact Nearest Neighbor (ENN)

javascript 
{
  $vectorSearch: {
    index: "vector_index",
    path: "embedding",
    queryVector: [...],
    exact: true,          // ENN mode
    limit: 50
  }
}

Spring Data MongoDB 4.5.0

java 
VectorIndex index = new VectorIndex("idx")
  .addVector("e", v -> v.dimensions(1536).similarity(COSINE))
  .addFilter("tenantId");

mongoTemplate.searchIndexOps(Movie.class).createIndex(index);

Use Cases

  • RAG systems - Hybrid retrieval ($vectorSearch + $search + filters)
  • Semantic recommendations - Movies, e-commerce products with vector similarity
  • Multi-tenant SaaS - ENN ensures exact retrieval per tenant after metadata filters
  • Real-time AI agents - Kafka → Atlas → Bedrock/LangGraph pipelines
  • Regulated search - Encrypted vectors + redaction filters for compliance

Pros & Cons

Advantages

  • Single datastore - Operational & vector data without sync tax
  • ANN + ENN + hybrid - All in one query language
  • Search Nodes - Workload isolation & cost control
  • Quantization + encryption - Secure, cost-efficient scale
  • Unified platform - Store, process, and search all data types

Disadvantages

  • Not in Community Edition - Vector features planned late 2025
  • No native distributed HNSW - Single cluster per region clustering
  • ENN latency scaling - Grows linearly, careful sizing needed
  • Maturing ecosystem - Tooling vs. dedicated vector DBs still developing

Future Outlook & Integrations

  • Community Edition Support [Late 2025]: Vector support for Community Edition targeted for late 2025
  • Voyage AI Integration [H2 2025]: Next-gen embeddings & retrieval metrics integration
  • Agentic RAG Blueprints [H2 2025]: LangGraph & CrewAI blueprints for agentic RAG workflows
  • Quantization Improvements [Ongoing]: Continuous improvements with auto-tuning and 4-bit quantization
  • Multi-Region Active-Active [Under Exploration]: Multi-region active-active vectors under exploration