From the AI Jungle to RAG in a Tab

“RAG”, “local AI”, “vector databases”… it’s easy to treat them as a checklist. I did too—until I tried to build a RAG app that runs entirely inside the browser

In this talk, I will show how a local-first RAG workflow fits together: ingest documents on-device, turn them into a searchable memory, retrieve the right context, and generate answers—without sending private data to a backend. The payoff is twofold: privacy by default and a new way to think about cost—shifting parts of the workload from centralized infrastructure to user devices..

Then we tackle the practical question: when is a local approach enough, when do you need a real vector index, and when does it cross the line where you need a server? We’ll use a simple scaling intuition (how retrieval work grows with corpus size and embedding dimension) to make that decision predictable, not magical.

Finally, I’ll turn RAG into something you can see that shows where queries land and why the retriever picked what it picked—so you can debug retrieval instead of guessing.