Do I need a vector DB if I'm only storing a few thousand embeddings?

Not necessarily.

If you’re only storing a few thousand embeddings, a full vector database is often overkill. In many cases, a simpler setup is enough:

Good alternatives for small scale

• In-memory search with NumPy / PyTorch if everything fits comfortably in RAM
• FAISS if you want fast similarity search without running a database
• SQLite/Postgres + embeddings in rows if you also need metadata and basic filtering
• Simple brute-force cosine similarity if the dataset is tiny and latency requirements are modest

When a vector DB starts to make sense

Consider a vector DB if you need:

• Millions of vectors
• Frequent inserts/deletes/updates
• Metadata filtering at scale
• Distributed or persistent serving
• Low-latency approximate nearest-neighbor search
• Multi-user / production-grade retrieval infrastructure

Rule of thumb

For a few thousand embeddings, you can usually start with:

• brute-force similarity search, or
• FAISS, or
• Postgres/SQLite if you need persistence and metadata

Then move to a vector DB later if the system grows or operational needs become more complex.

If you want, I can suggest the simplest option based on your stack (Python, JS, Postgres, etc.).