Definition
Plain language
A system that lets a language model itself search through a giant pile of documents, instead of relying on a separate search tool.
As stated in the literature
A retrieval framework built on a small Qwen3 model that wraps each document with a random code, processes the corpus once into a reusable cache with intra-document attention, and answers by generating the relevant document's code; trained with randomized IDs and an on-policy loss to generalize far beyond its native context length.
Why it matters: It lets a small model search through far more text than it could normally hold at once, offering an alternative to bolting on a separate retrieval tool.
For example, asked a question about a huge pile of documents, the model itself points to the right one by generating its assigned code, rather than calling out to a separate search engine.
Heard on the show
“They call their system BlockSearch.”Episode 198 — The Model That Knows the Answer and Can't Say It