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.
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.
We built a retrieval system over 100 million web documents. The biggest lesson was not about any single configuration. When every experiment takes 30 hours, the speed of your reindex-evaluate loop determines how good your system can get.
Context windows can hold more than ever, but more context doesn't mean better answers. Here's why retrieval still matters, especially for agents.
Search has a new user: agents. Longer, programmatic queries change the performance math of keyword retrieval. Here we take a closer look at Block-Max WAND when querying is driven by agents.
In agentic systems, retrieval is context injection. Better retrievers produce more convincing distractors; better reasoners trust those distractors more deeply. This is mutually assured distraction.
