Skip to main content
Welcome. This lesson surveys AWS options for storing and managing vector embeddings — numerical representations of text, images, and other data that capture semantic relationships. Embeddings enable fast, semantically meaningful retrieval for generative AI, semantic search, recommendations, and other ML-driven applications. Below you’ll find a concise review of relevant AWS services, their strengths, and common use cases.
A presentation slide titled "AWS Vector Database Services — Introduction" with three numbered panels. Each panel states that vector databases store numerical embeddings for text and images, embeddings enable fast semantically relevant searches for AI/ML, and AWS services help store and manage those embeddings.

​
AWS options for vector embeddings

​
Amazon OpenSearch Service (vector search for generative AI)

Amazon OpenSearch Service (the managed successor to Elasticsearch on AWS) supports high-performance k-NN vector similarity search and can act as a vector store for generative AI workloads.
  • High-performance k-NN search for fast semantic retrieval.
  • Real-time ingestion and indexing so newly generated embeddings become searchable immediately.
  • Integrates with ML tooling such as Amazon Bedrock and AWS SageMaker for embedding generation and model inference.
  • Serverless and managed deployment options allow auto-scaling to meet dynamic workloads.
Use cases: recommendation engines, multimodal (text + image) search, conversational AI with contextual retrieval, log analytics, and website search.
The image is a promotional slide titled "Amazon OpenSearch Service for Generative AI" showing a purple search/icon graphic and several feature boxes. The features mention managed vector-database service, high-performance vector similarity search, and scalable infrastructure for generative AI.
Key OpenSearch features for vectors:
  • k-NN vector search to find semantically similar items.
  • Easy integration with Bedrock and SageMaker embeddings.
  • Real-time ingestion and indexing to support interactive applications.

​
Amazon Aurora (PostgreSQL-compatible) & Amazon RDS for PostgreSQL (pgvector)

Both Amazon Aurora PostgreSQL-Compatible Edition and Amazon RDS for PostgreSQL support the pgvector extension, letting you store and query vector embeddings directly inside a relational database.
  • Combine SQL capabilities (transactions, joins, indexes) with vector similarity search using pgvector.
  • Store ML-generated embeddings from Amazon Bedrock, AWS SageMaker, or other LLMs and run vector queries alongside standard SQL operations.
  • Good fit for catalogs, recommendation systems, and applications that need ACID guarantees plus semantic search.
A presentation slide announcing that Amazon Aurora PostgreSQL‑Compatible Edition and Amazon RDS for PostgreSQL now support the pgvector extension. It notes pgvector is available on those services, enables storage and similarity searches with ML‑generated embeddings, and that embeddings capture semantic meaning from text processed by LLMs.

​
Amazon Neptune ML (graph + Graph Neural Networks)

Amazon Neptune ML extends Neptune with Graph Neural Network (GNN) capabilities for graph-native ML tasks.
  • Leverages frameworks like the Deep Graph Library (DGL) to simplify model selection and training for graph ML.
  • Ideal when relationships and topology matter: social graphs, knowledge graphs, fraud detection, or complex recommendations.
  • GNNs improve predictions by combining node/edge features with graph structure.
A presentation slide titled "Amazon Neptune ML" with a purple icon showing a database connected to a graph. Two caption boxes note that Graph Neural Networks (GNNs) enhance predictions using complex graph relationships and that it leverages the Deep Graph Library (DGL) for model selection and training.
Amazon MemoryDB is a Redis-compatible, in-memory data store suitable for ultra-low-latency vector retrieval.
  • In-memory storage gives microsecond reads and single-digit millisecond writes.
  • Scales to handle millions of vectors and high QPS workloads.
  • Can be configured for Multi-AZ durability for production reliability.
Use MemoryDB for real-time personalization, low-latency recommendations, and high-throughput embedding lookups.
A slide titled "Vector Search for Amazon MemoryDB" with five rounded boxes listing key features. The boxes highlight in-memory Redis-compatible design, microsecond reads/single-digit ms writes, support for millions of ML vectors with fast queries/updates, high QPS with >99% recall, and multi‑AZ durability.
Amazon DocumentDB (MongoDB-compatible) adds vector search to a managed JSON document database.
  • Keeps the flexible document model (JSON) while enabling millisecond-range semantic retrieval.
  • Works well for content-heavy or document-centric applications that combine full-text/document queries with semantic similarity.
  • Integrates with Bedrock, SageMaker, and third-party models for embedding generation and retrieval.
A presentation slide titled "Amazon DocumentDB (with MongoDB Compatibility) Support Vector Search" with a purple database icon. It lists key features: fully managed JSON document DB with MongoDB compatibility, vector search with millisecond response times, integration with Amazon Bedrock/SageMaker and third‑party ML models, and support for interactive log analytics and real‑time monitoring/website search.

​
Amazon Bedrock — custom knowledge bases and RAG

Amazon Bedrock supports retrieval-augmented generation (RAG) workflows and tooling for creating custom knowledge bases.
  • Store domain-specific documents and embeddings so RAG pipelines can retrieve relevant context at query time.
  • Useful to keep model outputs domain-aware and current without retraining base models.
  • Knowledge bases are updatable, improving recency and answer quality for generative apps.
RAG is widely used to fetch up-to-date or domain-specific information when answering user queries, and to provide context that reduces hallucination.
Exam tip: Memorize the AWS services that support vector embeddings and vector search — OpenSearch, Aurora/RDS PostgreSQL with pgvector, Neptune ML, MemoryDB, DocumentDB, and Bedrock (custom knowledge bases + RAG).

​
Quick comparison

​
Choosing the right vector store

Consider these factors when selecting a vector store:
  • Latency requirements: real-time interactive apps often require in-memory or optimized search stores.
  • Scale: expected vector count and QPS will influence managed service choice and architecture.
  • Data model: relational joins, document structure, or graph relationships favor different services.
  • Consistency/transactions: if ACID semantics are required, prefer Aurora/RDS.
  • Integration with ML tooling: verify integrations with Bedrock, SageMaker, or your embedding pipeline.

​
Summary

Vector embeddings are foundational for semantic search, multimodal retrieval, recommendations, and conversational AI. AWS provides multiple managed services to store, index, and query embeddings — choose based on latency, scale, data model, and integration needs:
  • OpenSearch: managed, scalable vector search with real-time ingestion.
  • Aurora/RDS PostgreSQL + pgvector: SQL + vector search.
  • Neptune ML: graph-native GNNs for relationship-rich data.
  • MemoryDB: in-memory, ultra-low-latency retrieval.
  • DocumentDB: document-mode vector search for JSON-centric apps.
  • Bedrock: RAG workflows and custom knowledge bases for generative AI.
Make sure you can explain trade-offs (latency, scalability, integration, and data model) when selecting a vector store, especially for exam and architecture design questions.

Watch Video