Nan Duan

A lot of real-world queries are bad queries — vague, missing context, phrased in ways the retriever cannot match against documents that would actually answer them. Duan's "Query Rewriting in RAG" line (2023, senior author) trains an LLM to rewrite the input query into a form the retriever can work with, and to do so adaptively based on what the system already has in context. The methodological consequence for ai100 is direct: when you measure a model's brand visibility, the upstream rewrite step is silently deciding what brands the retriever even sees as candidates — and that step is rarely audited.

Worth following when

you want to know what happens between "user types a query" and "retriever looks for documents" — the layer where many brand-mention decisions are already made.

Topics

LLM-based query rewriting for RAG; conversational query reformulation; the silent layer between user input and retrieval.

Key works

"Query Rewriting in Retrieval-Augmented Large Language Models" (2023, senior author); foundational QA and multimodal-NLP work through the 2010s; ongoing foundation-model work in industrial research.