Similarity searchintermediate
Cosine similarity measures the angle between two embeddings; sorting your chunks by it is the entire retrieval half of RAG.
Make the thing say "I do not know" — then build the eval set that proves it still does after you change the prompt.
An unanswerable question is refused without a model call, and every citation resolves.
Published in full, before you start — every point is one of these and there is nothing else.
Cosine similarity measures the angle between two embeddings; sorting your chunks by it is the entire retrieval half of RAG.
Talking to Claude is an ordinary HTTP POST — a model id, a max token count, and a list of messages.
An AI answer with no sources cannot be checked, and an answer that cannot be checked cannot be trusted — showing which chunk each claim came from is the whole point of RAG.
A proxy in front of a paid API needs a rate limit or one user can spend your whole budget, and a cancellation token so a closed browser tab stops the work.
A healthcheck is how the platform knows to restart your container, and running as a non-root user is how a bug stops being a breach.