Vector stores for AI apps: OpenSearch, pgvector, S3
Lesson 5 of 5 in Ingesting and Storing Training Data.
GenAI applications retrieve by meaning: text (or images, or audio transcripts) becomes an Embedding vector, and "find related content" becomes a k-nearest-neighbor search over millions of vectors. A Vector database is wherever those vectors live with an index (typically HNSW) fast enough to answer in milliseconds, plus metadata filters to narrow the search.
The MLA-C02 exam does not ask you to benchmark engines — it asks you to match scale, latency, and cost posture to the store. Three AWS options cover the questions.
| Option | Scale & latency | Cost posture | Reach for it when |
|---|---|---|---|
OpenSearch Service / OpenSearch Serverless | Large scale, low-latency k-NN; hybrid keyword + vector search, rich filtering | A dedicated search estate (cluster or serverless capacity) — the full-featured, premium option | Production RAG at scale; it is also the default vector store Bedrock Knowledge Bases creates for you (Serverless) |
Aurora / RDS PostgreSQL + pgvector | Moderate scale at SQL-query latencies | Reuses the relational database you already run — no new system | Vectors belong next to application rows: one SQL query joins similarity search with relational filters |
Amazon S3 Vectors | Very large vector sets; subsecond queries — slower than hot stores, built for scale over speed | S3-native, designed to minimize the cost of storing and querying vectors | Huge or infrequently queried collections — archives, batch RAG, long-tail knowledge bases |
Multimodal data has a storage split the exam likes: originals stay in S3 (images, audio, video, documents — a lake is the only store that takes them all), embeddings go to the vector store with metadata pointing back to the source object, and audio is usually transcribed before embedding. Multimodal embedding models (for example Titan Multimodal Embeddings on Bedrock) put text and images into the same vector space, so one index can answer "find images like this caption."
Other engines exist (MemoryDB for ultra-low latency, DocumentDB, Neptune for graphs+vectors) — the exam rewards matching the three above, not cataloging every option.
Which store do Bedrock Knowledge Bases use by default?
The quick-create path provisions an OpenSearch Serverless vector index for you. You can bring your own instead — Aurora PostgreSQL (pgvector) and several partner stores are supported — which is exactly the kind of "default vs bring-your-own" distinction exam options exploit.
Is DynamoDB a vector database?
No — DynamoDB has no native vector index. Scenarios pairing "key-value lookups" with "semantic search" want DynamoDB plus a vector store, not one service doing both.
Do I always need a managed vector store for RAG?
No. A small corpus can live in an in-process index (e.g. FAISS) inside your container. Managed stores earn their keep with scale, metadata filtering, durability, and concurrent access.
Interactive sorting exercise: Match the workload to the vector store.
Key terms: Vector database, Embedding
Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.