Vector Search

Vectors are mathematical representations of data, expressed as arrays of numbers that capture specific features or properties. In machine learning and artificial intelligence, vector embeddings refer to these numeric representations, often used to encode complex data such as text, images, or audio into a format computers can understand and process. Embeddings are designed to retain relationships and patterns within the data, such as similarities or contextual meaning. For example, in natural language processing, words with similar meanings are represented by embeddings close to each other in vector space. This approach enables efficient search, recommendation systems, and other machine learning tasks.

Vector search allows searching through data based on semantic meaning captured in vectors, whereas full-text search allows for keyword search (i.e., based on the actual text and any defined synonym mappings).

Yes, MongoDB is a vector database. Atlas is a fully managed, modern, multi-cloud data platform with a rich array of capabilities that includes text or lexical and vector search. Rather than use a standalone or bolt-on vector database, the versatility of our platform empowers users to store their operational data, metadata, and vector embeddings and seamlessly use vector search for indexing, retrieval, and building performant generative AI applications.

KNN stands for "K-Nearest Neighbors," which is the algorithm frequently used to find vectors near one another.

ANN stands for "Approximate Nearest Neighbors" and it is an approach to finding similar vectors that trades accuracy in favor of performance. This is one of the core algorithms used to power vector search. Our algorithm for Approximate Nearest Neighbor search uses the Hierarchical Navigable Small World (HNSW) graph for efficient indexing and querying of millions of vectors.

ENN stands for “Exact Nearest Neighbors” and it is an approach to finding similar vectors that might trade some performance in favor of accuracy. This method returns the exact closest vectors to a query vector, with the number of vectors specified by the variable limit. Exact vector search (ENN) query execution can maintain sub-second latency for unfiltered queries up to 10,000 documents. It can also provide low-latency responses for highly selective filters that restrict a broad set of documents into 10,000 documents or fewer, ordered by vector relevance.

$vectorSearch is an aggregation stage in MongoDB Atlas that lets you execute an Approximate Nearest Neighbor (ANN) or Exact Nearest Neighbor (ENN) query with MongoDB Query API filtering (e.g., “$eq” or “$gte”). This stage is supported on Atlas clusters version 6.0 and higher. The Atlas Search knnVector field type and knnBeta operator in $search are now deprecated.

MongoDB Vector Search supports embeddings from any provider that is under the 4096-dimension limit on the service.

We support the ingestion, indexing, and querying of scalar and binary quantized vectors from embedding providers. We also provide the option to implement automatic scalar and binary quantization of full-fidelity vectors in MongoDB.

Yes, MongoDB Vector Search can query any kind of data that can be turned into a vector embedding. One of the benefits of the document model is that you can store your embeddings right alongside your rich data in your documents.