Please note: This PhD seminar will take place online.

Dongfu Jiang, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Wenhu Chen

Modern retrieval systems typically expose a corpus through a fixed top-k similarity interface: a retriever first selects a small set of documents, and an agent or language model then reasons over the retrieved evidence. While this interface is efficient, it can become a bottleneck for agentic search tasks that require exact lexical constraints, sparse clue conjunctions, local context verification, multi-step hypothesis refinement, and the ability to recover evidence that may be filtered out too early. 

In this talk, I will present Direct Corpus Interaction (DCI), an alternative paradigm where an agent directly interacts with the raw corpus using general-purpose terminal tools such as grep, file reads, shell commands, and lightweight scripts, without relying on an embedding model, vector index, or retrieval API. I will discuss why conventional semantic similarity can be insufficient for agentic search, how DCI changes the interface between language agents and corpora, and how this simple setup performs across information retrieval benchmarks and end-to-end agentic search tasks. More broadly, the talk argues that as language agents become stronger, retrieval quality depends not only on the retriever or the model, but also on the resolution and flexibility of the corpus interaction interface.

Attend this PhD seminar virtually on Zoom.