# Where Vectors Live — Embeddings & Vector Search

Source: https://www.geekswithgeeks.com/en/embeddings/p-store

> Compare libraries, extensions to databases you already run, and dedicated vector databases.

## Library, extension or database

Three broad choices. (1) An **in-process library** such as FAISS: very fast, great for experiments and read-mostly data, but you handle persistence, replication, filtering and updates yourself. (2) A **vector extension for a database you already run** (for example pgvector for PostgreSQL, or vector search in Elasticsearch/OpenSearch, MongoDB or Redis): vectors sit next to your relational data and metadata, you reuse backups, access control and transactions, and filtering with SQL is natural; performance is good to large scale but may trail specialised engines. (3) A **dedicated vector database** (such as Qdrant, Milvus, Weaviate, Pinecone): built for scale, filtering, hybrid search, replication and operations, at the cost of another system to run or pay for. Start with what you already operate; move to a specialised engine when measured needs (scale, latency, features) demand it. The next course in this series compares vector databases in detail.

## Store, update, scale, protect

Choose where vectors live, keep them fresh, control cost and protect the data they came from.

![Four concerns: storage, freshness, scale, security.](assets/figures/embeddings/section-6-map.svg) — Figure 6.1 — Storage, freshness, scale and security.

## Where to start

A rough guide; measure before you commit.

```text
Situation                                        Reasonable start
experiment / notebook / offline batch              FAISS or numpy (flat)
already on PostgreSQL, < ~ tens of millions        pgvector (+ SQL filters)
already on Elasticsearch/OpenSearch, want hybrid   its vector + BM25 search
large scale, heavy filtering, many tenants         dedicated vector database
unsure                                             start simple; keep an abstraction so you can move
```

## Hide the store behind an interface

A small wrapper with add, delete and search lets you change engines later without rewriting the app.

**Quiz:** What is an advantage of a vector extension in a database you already run?

- [ ] It is always faster than every other option
- [x] Vectors sit with your data, reusing backups, access control and SQL filters
- [ ] It needs no storage
- [ ] It removes the need for embeddings

*Answer:* Vectors sit with your data, reusing backups, access control and SQL filters. One system to operate often beats two, until scale demands otherwise.