Key takeaways
- AWS offers OpenSearch, Aurora with pgvector, MemoryDB, DocumentDB and S3 Vectors for vectors.
- Dedicated vector databases are usually NoSQL, while relational databases can add vector support.
- LlamaIndex is a framework that connects to vector databases rather than being one.
- Vector databases store embeddings, IDs, metadata, text and indexes in memory, disk or object storage.
- Vector search is becoming a standard database feature, so choose the simplest option that works.
A vector database stores embeddings with their metadata and finds the most similar ones quickly, but the term covers dedicated systems, cloud services and features inside ordinary databases. I build retrieval systems such as RAG.NextUpgrad, which combines dense vectors with BM25 and reranking, so this page covers AWS options, whether vector databases are relational, SQL or NoSQL, LlamaIndex, what they store and where, setting one up, building one, learning the topic and whether we are at peak vector database.
Does AWS have a vector database?
Yes, several options. Amazon OpenSearch Service offers vector search, including a serverless vector engine; Aurora PostgreSQL and RDS support pgvector; MemoryDB and DocumentDB add vector search; and Amazon S3 Vectors stores vectors cheaply in S3.
Also asked as: does aws has vector database
Amazon Bedrock Knowledge Bases provides managed RAG on top of these stores, handling chunking, embedding and retrieval. Choose by existing stack and scale: teams on PostgreSQL often use Aurora with pgvector, search-heavy applications suit OpenSearch, low-latency in-memory needs suit MemoryDB, and very large, cost-sensitive archives suit S3 Vectors. Third-party databases such as Pinecone, Qdrant and Milvus also run on AWS through marketplaces or self-hosting. Check each service's current regions and features before designing.
Is a vector database relational?
Dedicated vector databases are generally not relational. They store vectors with metadata in collections and focus on similarity search rather than tables, joins and foreign keys. However, relational databases can add vector features, such as PostgreSQL with pgvector.
Also asked as: is vector database relational · how to setup vector database · how to learn vector database
This gives two styles. A dedicated vector database keeps embeddings separate from business data, syncing through pipelines. A relational database with vector support keeps everything together, so one query can join users, permissions and document vectors. For many applications, the relational approach simplifies architecture and consistency. Dedicated systems can perform better at very large scale or with heavy filtering, so test your workload before splitting data.
Is a vector database SQL or NoSQL?
Most dedicated vector databases are NoSQL systems with their own APIs and query languages, such as Qdrant, Pinecone, Milvus and Weaviate. Some systems offer SQL access, including PostgreSQL with pgvector, ClickHouse and several cloud data warehouses.
Also asked as: is vector database sql or nosql
Vector search is better seen as a capability than a database category. Document databases such as MongoDB Atlas, search engines such as Elasticsearch, key-value stores such as Redis, and SQL databases all now provide it. Choose based on how the rest of your data is stored and queried. If your team knows SQL and data is relational, SQL with vectors is often simplest; dedicated NoSQL vector databases suit specialised, large-scale retrieval.
Is LlamaIndex a vector database?
No. LlamaIndex is a data framework for building LLM applications, handling document loading, parsing, chunking, indexing, retrieval and agents. It connects to vector databases rather than being one, though it includes a simple in-memory vector store for prototypes.
LlamaIndex supports dozens of vector store integrations, such as Qdrant, Chroma, pgvector and Pinecone, letting you switch backends with little code change. It also offers advanced retrieval patterns, such as hierarchical and recursive retrieval, and document parsing tools. LlamaCloud, its commercial service, adds managed parsing and indexing. Use LlamaIndex to organise the RAG pipeline and a vector database to store and search embeddings at scale.
What does a vector database store?
A vector database stores embeddings, which are lists of numbers representing meaning, along with an ID for each item and metadata, such as document title, source, date or permissions. Many also store the original text chunk.
Also asked as: what does vector database store
It also stores index structures, such as HNSW graphs, that make similarity search fast, and sometimes compressed or quantized copies of vectors to save memory. Metadata enables filtering, such as searching only a user's documents. Storing the text alongside vectors lets the application return passages directly for RAG, with sources for citations. Some systems store multiple vectors per item, such as dense and sparse vectors for hybrid search.
Where is a vector database stored?
It depends on the system: in memory for speed, on local disk for persistence, in cloud object storage for low-cost scale, or on a managed provider's infrastructure. Many combine memory caching with disk or object storage.
Also asked as: where is vector database stored
Local tools such as Chroma and LanceDB store files in a folder on your machine. Self-hosted servers such as Qdrant or Milvus store data on the server's disks, often in containers with mounted volumes. Managed services, such as Pinecone or cloud provider offerings, store data in their data centres in regions you choose. For sensitive data, check where data is physically stored, how it is encrypted and how backups work.
How do you set up a vector database?
For a quick start, run Qdrant with one Docker command, install Chroma with pip for local Python use, or enable pgvector in PostgreSQL with CREATE EXTENSION vector. Then create a collection or table with your embedding dimension and distance metric.
Next, embed documents with your chosen model, insert vectors with IDs, text and metadata, and run a similarity query to confirm results. Create an index such as HNSW for speed as data grows. Test with a few hundred documents before loading everything. For production, configure persistence, backups, authentication, monitoring and resource limits, or use a managed service. Keep the embedding model name and version recorded with your data, since changing models requires re-embedding everything.
How do you build a vector database?
To learn the idea, store vectors in an array, compute cosine similarity between a query and every vector, and return the top results. Then add an approximate index, such as HNSW through hnswlib or FAISS, persistence to disk and metadata filtering.
Also asked as: how to build a vector database
Building a toy vector database teaches why indexes matter: brute-force search becomes slow as vectors grow, while HNSW finds neighbours in a fraction of the time with small accuracy loss. Add an API with FastAPI for inserting and searching, and save index files so data survives restarts. Production databases add replication, sharding, concurrent updates and crash recovery, so use an existing database for real applications.
How do you learn vector databases?
Learn embeddings and similarity metrics first, then approximate nearest neighbour indexes such as HNSW, metadata filtering, hybrid search and evaluation. Free resources include DeepLearning.AI's short course on vector databases, Qdrant and Weaviate documentation, and Pinecone's learning articles.
Practise by building a semantic search over your own notes: embed documents, store them in Chroma or pgvector, and query by meaning. Then build a small RAG app that answers questions with citations. Compare two databases on the same data, measuring recall and latency. Reading the HNSW paper deepens understanding. Combine database knowledge with chunking, embeddings and reranking, since retrieval quality depends on all of them. Write a short post about what you learned to consolidate it.
Are we at peak vector database?
Possibly for standalone hype, not for vector search itself. Vector search is becoming a standard feature of general databases, which reduces the need for separate systems in many projects, while retrieval remains essential for AI applications.
Long context windows let some applications skip retrieval, and many teams use pgvector or existing search engines instead of new databases. Dedicated vector databases are responding with stronger hybrid search, filtering, scale and cost efficiency, and some consolidation through acquisitions is likely. For builders, the lesson is to treat vector search as one component, choose the simplest option that meets requirements, and keep retrieval design portable. Retrieval quality, not the database brand, decides answer quality.
Sources
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
