Thinking traces
Code mode for agentic retrieval
Frontier models are strong at code, and code lets agents manage retrieval without filling context. The next step is programmable retrieval primitives, not just programmable query loops.
Blog
Proof
Agent queries cost more to serve
A human types two words; an agent sends ten. That shift in query mix means 8.4x the cost to serve.
Relevance
This is what agentic retrieval looks like
GPT-5 searches like a power user. Phrase quotes in 98% of sessions. Median query length past the human 99th percentile (AOL query log). The new user of search.
Proof
What we learned building a 100M-document search engine, part 3: Hybrid search depends on query type
Hybrid search does not have one best weight. Agentic, all-words-included, and human-style queries each stress retrieval differently, which means evaluation has to be stratified.
Proof
What we learned building a 100M-document search engine, part 2: ANN Tuning
We expected ANN tuning to be straightforward. It was not. Embedding bias, graph connectivity, and quantization loss interacted in ways that only surfaced at 100 million documents.
Relevance
Deep research is a retrieval problem
Give the agent the relevant documents and it answers 93% of questions correctly. Make it find those documents with a weak retriever and it scores 14%. In BrowseComp-Plus, that gap makes retrieval impossible to ignore.