MongoDB Atlas Vector Search

Developer Documentation & Technical Reference

Index Creation

Define a Vector Search index using the Atlas UI, Atlas CLI, Atlas Admin API, or the MongoDB drivers. Indexes support HNSW (Approximate Nearest Neighbor) and Flat (Exact Nearest Neighbor) configurations. Example index definition:

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

Querying with $vectorSearch

Use the $vectorSearch aggregation stage to perform ANN or ENN queries:

db.collection.aggregate([
  {
    $vectorSearch: {
      index: "vector_index",
      path: "embedding",
      queryVector: [...],
      numCandidates: 150,
      limit: 10,
      filter: { category: "technology" }
    }
  }
])

Key Technical Capabilities

• Hybrid search: Combine $vectorSearch with $search (lexical/BM25) using $unionWith or the $rankFusion operator (Reciprocal Rank Fusion).
• Metadata filtering: Pre-filter or post-filter results using any MongoDB query expression.
• Search Nodes: Isolate vector workloads on dedicated compute nodes, scaling independently from the primary cluster.
• Automated Embeddings: Configure embedding generation via Voyage AI models directly in the Atlas UI — embeddings are created and updated automatically on insert/update.
• Multitenant support: Flat Indexes simplify multitenant vector search architectures with efficient per-tenant filtering.
• Driver support: Available across all official MongoDB drivers (Python, Node.js, Java, Go, C#, etc.).