Researchers led by Nabaraj Subedi examined retrieval‑augmented generation (RAG) on growing document sets. In a Wyoming DOT system, expanding the corpus from 54 to 1,128 documents cut precision@10 by 63% and lowered recall of the answer‑bearing document by 23%. The paper proposes a soft‑scoping technique that limits retrieval to relevant domains without extra training. The authors recommend applying soft scoping first, then using a single synthesis call.
Why it matters
For developers of RAG‑based chatbots, adding more documents can actually make answers less accurate unless domain scoping is used.
