AWS vector solutions: Build agentic AI where your data lives

Agentic AI is changing how you work, and vector search powers the retrieval layer that makes agents accurate, contextual, and grounded in real data. Agents plan, reason, and take action across multi-step workflows, making fast, relevant access to your organization’s knowledge essential.

That knowledge already has a home across databases, object stores, search engines, and unstructured sources such as PDFs, recorded video calls, and the systems your teams use every day. AWS vector solutions bring intelligent search and retrieval to your data where it already lives, helping agents find and use the right context without requiring you to move or duplicate your data.

For new workloads where no existing data store applies, we offer a clear decision model across six purpose-built solutions so you can choose the right vector solution for your agentic AI and analytics workloads.

Why vectors matter and top use cases for agentic AI

Vectors are the language of AI. They bridge frontier models and the scattered organizational knowledge accumulated over decades. By representing data as high-dimensional vectors, applications can understand semantic meaning, identify relationships across text, images, audio, and video, and maintain context across sessions. Whether you’re working with product descriptions, security logs, or media libraries, vectors convert everything into a shared mathematical space so you can compare and search across modalities.

Combined with frontier AI models, vectors are the foundation for intelligent, context-aware, personalized, and user-specific experiences:

Retrieval Augmented Generation (RAG) and knowledge bases ground frontier model responses with trusted data retrieved at runtime, improving accuracy, reducing hallucinations, and generating responses aligned with organizational knowledge.

Semantic search retrieves information based on meaning and intent rather than exact keyword matches, allowing users to discover relevant content even when different terminology is used.

Hybrid search combines lexical search with semantic search to deliver comprehensive results across structured and unstructured data.

GraphRAG combines semantic search with knowledge graphs to deliver accurate, context-rich, and traceable responses for enterprise scenarios requiring multi-step reasoning.

Knowledge graphs connect entities, such as people, products, documents, and concepts through explicit relationships, supporting more intelligent search, discovery, and AI-powered reasoning.

With the options outlined in the previous section, vectors support several use cases, including:

Real-time recommendation systems identify products, content, or experiences that align with user interests for personalization through vector similarity across retail, media, and entertainment.

Anomaly and fraud detection identifies unusual patterns in high-dimensional data, supporting earlier detection of threats, cyber security issues, or equipment failures.

Multimodal content discovery searches across text, images, audio, and video using a single query based on semantic meaning rather than file type or metadata.

The good news is you don’t need a new database for most of these use cases. With AWS, you get vector capabilities where your data already lives, across the services and databases that you already know and use, with no data migration required.

The following figure shows the breadth of AWS vector capabilities and top use cases for agentic AI.

Figure 1: Breadth of AWS vector capabilities for your agentic AI use cases

First principle: Add vectors where your data already lives

This is the principle that guides our entire approach. If you already have an AWS data store, add vector search to that data store. Don’t introduce a new service unless there’s a compelling reason to do so. Vectors stay with the source data, removing cross-service hops and combining vector search with native query capabilities.

When you use your existing data store, you remove the learning curve for new programming tools, APIs, SDKs, and more. You also can be confident that your existing data stores meet your requirements. As an example, your databases proven in production for scalability, availability, and performance will continue to deliver now with vector search. Finally, when your vectors and data are stored in the same place, your applications run faster. There’s no data sync or data movement to worry about. You also realize cost savings by building on investments that you’ve already made.

If your data is already in Amazon OpenSearch Service, Amazon Simple Storage Service (Amazon S3), Amazon Aurora PostgreSQL, Amazon DynamoDB, Amazon ElastiCache for Valkey, or Amazon Neptune, add vectors where the data already is. The right vector search solution follows the data, not the other way around.

For new workloads, identify your dominant requirement: latency, cost, or access pattern, and choose the engine optimized for it. Many workloads need a balance of search, scale, and agentic AI integration. For those, default to Amazon OpenSearch Service, which combines lexical, vector, hybrid, and agentic search in a single system with high throughput, low latency, and relevant results at scale.

The following decision model can help you select the right vector solution based on your workload requirements.

Figure 2: Decision model for vector engines

Amazon OpenSearch Service: The default for new workloads

Amazon OpenSearch Service is a managed retrieval engine that combines lexical, vector, and hybrid search in a single system with high throughput, low latency, and relevant results at scale. It supports multiple indexing strategies, vector quantization, and metadata filtering, scaling from simple RAG applications to advanced multi-signal retrieval. Machine learning (ML)-powered auto-optimization removes manual tuning by selecting the right configurations automatically. GPU acceleration indexes massive datasets up to 10x faster at a quarter of the cost, while UltraWarm and Writable Warm tiers reduce storage costs for less frequently accessed data.

Choose OpenSearch Service as the default for new workloads because most new workloads don’t have a single dominant requirement. They need a balance of search, scale, and agentic AI integration. OpenSearch Service provides the most flexibility across latency, vector volume, queries per second (QPS), cost efficiency, hybrid search, and ease of adoption. It covers the broadest set of use cases including RAG, anomaly detection, multimodal content discovery, and any workload requiring hybrid search. It supports multi-billion vector volumes, handles thousands of QPS, and serves more than 100,000 monthly active customers processing over 10 trillion requests per month.

The next generation of Amazon OpenSearch Serverless is built for agentic AI and dynamic workloads. It autoscales 20x faster than its previous generation, provisions in seconds, and ramps from zero to thousands of requests per second. It also scales back to zero when idle, delivering up to 60% cost savings compared to provisioning for peak capacity. You only pay for consumed capacity. If your agents aren’t running, you pay nothing.

Adobe adopted OpenSearch Service to scale its Acrobat AI Assistant to serve hundreds of millions of users. This is a conversational generative AI engine integrated directly into Adobe’s document ecosystem.

Amazon S3 Vectors: Cost-optimized vector storage at any scale

Amazon S3 Vectors, a capability of Amazon S3, is the first cloud object store with native support to store and query vectors. It brings the cost structure, scale, and simplicity of S3 to vector storage, reducing the cost of uploading, storing, and querying vectors by up to 90 percent compared to specialized vector databases. This makes it cost-effective to build and maintain billion-scale vector indexes that improve AI agent memory, context, and semantic search across content stored in Amazon S3 with zero infrastructure to manage.

Since the general availability of S3 Vectors, customers have been performing tens of millions of queries per day on average, a more than 5x increase over the preview period. Two recent enhancements improve the query experience and pricing. First, S3 Vectors now supports up to 10,000 search results per query, a 100x increase that’s especially valuable for multi-stage retrieval pipelines that apply reranking, aggregation, or deduplication to produce a more relevant result set. Second, query charges on vector indexes with over 10 million vectors are now reduced by up to 80%, significantly lowering costs for running similarity search across large-scale AI, RAG, and semantic search workloads.

Choose Amazon S3 Vectors when you need cost-effective vector storage with simple vector search and metadata filtering for infrequent queries of large vector datasets. It’s recommended when latency can be approximately 100 ms or more, or for fast-growing vector datasets at moderate QPS. It supports up to two billion vectors per vector index with pay-per-query pricing, so you pay for stored vectors while query costs accrue only when you search. Common use cases include semantic search over data lakes, RAG-based knowledge retrieval, large-scale vector storage, and batch retrieval. Customers also work directly with vectors and indexes through the S3 API.

BMW Group uses S3 Vectors as a building block for its hybrid search solution that is powered by an intelligent search agent built with Amazon Bedrock AgentCore. Engineers can query 20 petabytes of data in plain natural language, combining S3 Vectors for semantic similarity searches and Amazon Athena for SQL queries.

Amazon DynamoDB: Single-digit millisecond vector search at any scale

Amazon DynamoDB is a serverless, fully managed, distributed NoSQL database with single-digit millisecond performance at any scale. Vector search with DynamoDB delivers single-digit millisecond latency at over 99 percent recall designed for any scale, even trillions of vectors. It’s fully serverless with no servers to provision, patch, or manage, and no software to install, maintain, or operate. The zero infrastructure management that you love with DynamoDB, including no versions, no maintenance windows, and no downtime maintenance, is also available on vector search. Vector search in DynamoDB introduces a new index you create on the attribute that stores vector embeddings. It supports up to 4,096 dimensions, Euclidean, cosine, and dot-product distance functions, and inline filtering. DynamoDB vector search works with DynamoDB global tables, both with multi-Region eventual consistency and strong consistency.

DynamoDB serves over a million customers. Customers already use DynamoDB today for agentic workloads, such as holding session context during conversations and tracking state across multi-step tasks. With vector search, DynamoDB can support semantic retrieval for long-term memory in agentic applications. Other use cases include RAG, multimodal similarity search, and recommendation engines. And you can do all of this in one serverless database, no separate vector store to maintain and synchronize data, and no new API to learn.

Choose DynamoDB vector search when your operational data already lives in DynamoDB, or you need single-digit millisecond vector search at any scale with zero infrastructure management.

Globant is a digitally native company that builds AI-powered products and digital transformation solutions for enterprises around the world.

“We already build client solutions on DynamoDB, so having native vector search in the same database is extremely valuable — no need to replicate data into a separate vector store or manage a second system. It’s fully serverless, scaling automatically across virtually any scale, so we only pay for what each client uses, and it delivers the real-time, low-latency search that user-facing AI experiences demand. It lets our engineers focus on building AI apps instead of managing infrastructure.”

— Gastón Milano, CEO Enterprise AI, Globant

Amazon ElastiCache for Valkey: Microsecond latency for semantic caching

Amazon ElastiCache is a serverless, fully managed caching service delivering microsecond latency performance with full Valkey, Memcached, and Redis OSS compatibility.

With ElastiCache for Valkey, you can build more personalized, context-aware responses by implementing memory mechanisms that surface cross-session conversation history to large language models (LLMs). Similar to how a cache reduces database costs and improves application performance, semantic caching reduces the cost and latency of using LLMs by serving cached responses for semantically similar prompts. You can also use vector search to power RAG on large datasets to improve response relevance and reduce hallucinations by grounding outputs with real-world data.

Choose ElastiCache for Valkey when the workload has a microsecond latency requirement. This includes real-time recommendation engines, session-based personalization, latency-critical RAG pipelines, and semantic caching. It supports up to one billion vectors for microsecond latency vector search.

Sanoma uses ElastiCache for Valkey to turn human moderator decisions into vectors that inform future AI moderation calls in real time, with no retraining required. Today, 30 percent of comments are matched against past decisions, and 6.5 percent receive a different, more accurate outcome as a result.

Amazon Neptune: For GraphRAG

Amazon Neptune is a serverless graph database service for connected data and improved AI accuracy.

When your workload involves highly connected data, or multi-hop reasoning, Amazon Neptune uniquely solves this by combining graph traversal with vector similarity in a single query.

Neptune delivers low-latency vector search with 2–3 billion vector capacity, and built-in multi-hop reasoning and traceability. That last point is critical. Regulated industries must show why a result was returned, not just what was returned. This transparency is what makes Neptune essential for compliance, risk, and security workloads.

Choose Neptune when the workload needs to combine graph traversal with vector similarity in a single query. With Neptune, you can traverse connections (for example, “which teams own the services affected by this alert?”) and combine that relationship context with vector similarity in a single query. Many of our customers use Neptune for workloads that need a knowledge graph, such as financial and compliance risk, pharmaceutical research, drug discovery, and security intelligence.

Deloitte uses Amazon Neptune with the AWS GraphRAG Toolkit to power a Security Intelligence Center that combines graph-based knowledge retrieval with generative AI. By connecting policy interpretation, operational enforcement, and real-time metrics through GraphRAG, Deloitte delivers predictive security guidance grounded in timely organizational context.

Amazon Aurora PostgreSQL: SQL-native vector search

Amazon Aurora delivers high performance and availability at global scale for PostgreSQL. You can turn on the pgvector extension and use optimized reads for high-performance vector search.

Amazon Aurora PostgreSQL with pgvector 0.8.0 delivers 9x faster indexing and 100x more relevant filtered results. This is a major leap. It combines vector search with the full SQL query surface: joins, aggregations, WHERE clauses, and ACID transactions in a single engine.

Choose Aurora when your source data already lives in Aurora. It’s recommended when you want to use SQL, or when you need to consolidate vector and relational workloads into a single database. It delivers low latency and supports hundreds of billions of vectors, and is ideal for multi-tenant software as a service (SaaS) applications, agentic memory stores, and RAG with structured data context.

LeadSquared, a SaaS CRM platform serving sectors such as BFSI, healthcare, and education, built their solution using Aurora PostgreSQL with pgvector and Amazon Bedrock for RAG. They accelerated chatbot deployments by using Aurora for both vector embeddings and essential operational data.

Combining solutions for complex architectures

Some use cases benefit from combining multiple vector services. You might have a workload best served by an architecture that combines Amazon ElastiCache for Valkey for semantic caching, Amazon OpenSearch Service for hybrid search, and Amazon S3 Vectors for persistent vector storage. Together, this combination allows you to scale from zero to trillions of vectors while optimizing for latency and cost across hot, warm, and cold tiers.

When evaluating complex workloads with potentially competing requirements, consider three dimensions simultaneously:

Performance and scale: Latency profile, vector volume, QPS requirements, and algorithm tuning for accuracy.

Cost and operations: Cost-latency trade-offs, tiering needs, operational model preference (serverless or provisioned), and existing team expertise.

Search capabilities: Hybrid search, SQL support, complex metadata filtering, full-text search, and multi-vector search.

The following table can help you select the right vector engine so you can optimize for latency, cost, or access pattern requirements.

Service Latency (P95) Vector Volume QPS Key Differentiator

Amazon OpenSearch Service 10–100 ms Hundreds of billions Thousands

Lexical, semantic, hybrid, and agentic search.

Broadest capability set, including serverless.

Amazon S3 Vectors > 100 ms Multi-billion Low hundreds / index Pay-per-query, lowest cost for infrequent queries

Amazon DynamoDB < 10 ms Trillions Virtually unlimited. (Thousands per vector index partition) Single-digit millisecond latency vector search at any scale

Amazon ElastiCache for Valkey < 10 ms Over 1B vectors Thousands Microsecond latency, semantic caching

Amazon Aurora PostgreSQL 10–100 ms Hundreds of billions Thousands SQL-native relational + vector

Amazon Neptune 10–100 ms 2–3B vectors Hundreds to thousands Graph traversal + vector (GraphRAG)

Table 1: AWS vector solutions

Getting started

Get started today with AWS vector solutions for semantic search and self-managed knowledge bases and RAG. Start by adding a vector solution to where your data exists today. For new workloads, use the decision model to help select a vector solution based on your latency and scale requirements, or default to Amazon OpenSearch Service if you’re not sure. If you’re interested in a fully managed knowledge base and RAG solution, read this post on Amazon Bedrock Knowledge Bases.

Join our webinar, AWS Vector Solutions Overview: Intelligence for Agentic AI Where Your Data Already Lives, for a detailed look on choosing the right vector engine and the latest updates across our portfolio. Contact your AWS account team for a hands-on workshop or proof of concept.

About the authors