Generation endpoint with a structured prompt

POST /answer feeding retrieved chunks and the question into an LLM call using a structured prompt template - system instructions, a retrieved-context block, the question, and an explicit instruction to cite sources. Return the answer plus the source chunks used.

Returning the chunks alongside the answer is not a nicety. It is the seam that makes the next task possible, and it is what separates a system you can audit from one you must trust blindly.