interactive · 12 min

A Tiny Retriever

Objective: retrieve the most relevant document by simple word overlap, then answer only from it and cite its id.

Real systems swap word-overlap for embedding similarity (subsystem D), but the shape is identical: score, rank, answer only from the top chunk. The model — your MockLLM-style policy — never invents facts here; it is constrained to fetched context.

Best practice: return the source id with every grounded answer; an answer with no retrievable source is a hallucination risk to flag.

Next: RAG Failure Modes