Choosing the vector layer

Lesson 1 of 3 in Vector Stores at Professional Scale.

Under every RAG answer sits a vector store, and the exam tests whether you can pick one like an architect instead of a tourist. The managed path: Bedrock Knowledge Bases quick-creates an Amazon OpenSearch Serverless index and owns chunking, embedding, storage, and retrieval end to end. The bring-your-own path keeps the same Knowledge Bases front door but supplies the store: OpenSearch (Serverless or managed cluster — highest sustained-QPS serving, hybrid and neural search via the Neural plugin), Aurora PostgreSQL with pgvector (vectors beside relational data — SQL joins and filters in the same engine), Amazon S3 Vectors (dramatically cheaper storage for huge or cold corpora at modest query rates), or partner stores (Pinecone, Redis, MongoDB Atlas) where a contract already exists.

The AIP-C01 exam guide adds two professional wrinkles. Amazon RDS with S3 document repositories: the pattern where vectors and metadata live in the database while the authoritative documents stay in S3, joined by URI. And DynamoDB with vector databases: DynamoDB is not the vector index — it is the millisecond key-value companion holding per-document metadata, sync state, and session context beside whichever store holds the embeddings.

Pick the vector layer

Interactive decision tree — outcomes:

Knowledge Bases quick-create (OpenSearch Serverless)
Bedrock creates and manages the vector index and the whole ingest-retrieve loop. Exam cues: "minimal operational overhead", "no vector infrastructure to manage".
OpenSearch Service (Serverless or managed cluster)
The high-performance serving choice: k-NN at scale, neural and hybrid queries, sharding and multi-index control. Both deployment modes are supported Knowledge Bases backends.
Aurora PostgreSQL + pgvector
IVFFlat/HNSW indexes and distance operators inside PostgreSQL — one engine for similarity plus the relational joins and filters your application already needs.
Amazon S3 Vectors
Vector buckets with native indexes: sub-second queries at a fraction of dedicated-store cost. The fit is big, cost-sensitive, lower-QPS corpora — not hot-path serving at thousands of QPS.
Partner store (Pinecone, Redis, MongoDB Atlas)
Knowledge Bases connects to these too — legitimate when the estate already runs one; rarely the exam's "most operationally efficient" pick on AWS-native stems.

The vector layer, store by store
Store	Strengths	Watch out	Exam cue
Knowledge Bases quick-create (OpenSearch Serverless)	Fully managed ingest → chunk → embed → retrieve	Less control over engine internals	"Minimal operational overhead" + RAG
OpenSearch Service	High-QPS k-NN, hybrid/neural queries, sharding & multi-index control	You own sizing (cluster) or capacity-unit economics (serverless)	"High sustained query volume", "hybrid search"
Aurora PostgreSQL + pgvector	Similarity + SQL joins/filters in one engine; IVFFlat & HNSW	Vector QPS competes with relational load	"Combine vectors with relational/business data"
S3 Vectors	Cheapest at-rest vectors; elastic; sub-second queries	Not built for sustained thousands of QPS	"Massive corpus, cost-sensitive, moderate traffic"
DynamoDB (companion)	Millisecond metadata, sync state, session context beside the vectors	It is not the similarity index itself	"Metadata and embeddings" paired architecture

Three retrieval products the options love to swap

Amazon Kendra is managed enterprise search: 40+ built-in source connectors (SharePoint, Confluence, Salesforce…), natural-language queries, and ACL-aware results — each user sees only what their source permissions allow. "Search our wikis and SharePoint with permissions respected" → Kendra, no vector plumbing exposed, no generation to build. (Closed to new AWS customers since July 2026 in favor of Knowledge Bases, yet still on the exam's in-scope list.) Bedrock Knowledge Bases is managed RAG retrieval: it grounds a foundation model — you own the generation and the application. OpenSearch is the engine: you own the index, the query DSL, and the operations. Two eliminators ride along: Amazon Q Business is the finished RAG-assistant product — the moment a stem asks you to build anything, it is out — and Amazon Neptune is the graph store behind GraphRAG-style Knowledge Bases, for questions where relationships, not passages, carry the answer.

Interactive checkpoint quiz (4 questions) — open this page in a browser to take it.