Decide whether your RAG project needs a dedicated vector database

Every roundup insists you need a dedicated vector database. Your RAG project is still a folder of a few dozen files.

BeginnerRead 5 minDoing: half an hour to an hour, mostly arithmetic on your own numbers rather than new code6 steps

▶ Start

What you'll have at the end

A one-paragraph decision note naming your store choice and the two facts that drove it

Before you start

You need

A working script that already turns your own documents into chunks, embeds them, and searches an in-memory array or list with cosine similarity, the way a first RAG pipeline typically starts.

Not covered

Building the filter, hybrid search, or reranker a growing project eventually needs; this only covers picking and justifying where vectors live today.

Leans on

Find your real pressures

Count what you actually have
Time an exact search at that scale
Check whether writes are batch or stream
Check whether a filter would drop results

Turn it into a decision

Score the table against the field's line
Write the memo and set a tripwire

Where to go next

Ground RAG answers in your own documents

for the in-memory chunk-and-embed pipeline this recipe's own count and timing assume you already have

Metadata filters that stop your RAG search returning the wrong department's file

once your filtering row actually crosses its line, this builds the filter itself

Checked 18 Aug 2026

Part of the RAG (Retrieval Augmented Generation) cookbook