Today's concept: RAG architecture, retrieval augmented generation
Most explanations of RAG still picture a person typing a question into a search box. Two papers from the last few months say that picture is out of date, and it changes what you should be testing.
RAG, retrieval augmented generation, means a model looks things up before it answers. A retriever searches your documents, picks the passages that match, and hands them to the model as reading material.
The early version ran that search once, on the user's own words. In agentic RAG the model writes the searches itself: it rewrites the question, splits it into parts, and searches again as it learns more.
In September, a team at Perplexity AI published Q2D-Web, a benchmark built from nine months of their own traffic: about 70,000 agent-written queries over 190 million web documents, in ten languages. For every original question, the agent fired several supporting searches of its own. The authors suggest that embedding models trained mostly on human queries do worse on those machine-written ones.
In July, Agent Retrieval Bench looked at coding agents choosing which files in a repository to read. On 27 to 35 percent of the logged runs, the agent missed every file it actually needed.
What I take from both: if you evaluate your retriever with questions people wrote, you are testing traffic your system barely sees. The agent has become the main user of your search index, and it does not search the way we do.
Quick check before you scroll: In the Perplexity traffic behind Q2D-Web, roughly what share of the searches the agent ran were the user's original question?
Full breakdown + the answer: frankduah.me/learnings/2026-10-23-rag-architecture-retrieval-augmented-generation
If your network builds with AI, a repost puts this in front of people who don't follow me yet.
New here? I post a bite-size AI / ML concept like this every day. Follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH
#AI #LLM #AIAgents #MachineLearning
The answer
About one in five. The paper counts 17.7 percent of queries as primary and 82.3 percent as support queries the agent wrote itself, roughly four for every original question, so most of what the retriever serves is machine-written.