Seoul, KR · Sookmyung Women's University

Chaewon Sung

AI/ML researcher working on modular RAG pipeline diagnostics — bridging research and product engineering.

// locating which module actually bottlenecks the pipeline, not just the average gain

Research

KDD '26

Beyond Average Oracle Gains: Split-based Bottleneck Analysis for Modular RAG Pipelines

Split-based analysis that identifies which stage of a modular RAG pipeline — retrieval, reranking, or generation — is the true bottleneck, rather than relying on average oracle gain alone.

with Yoonhyuk Choi · AI Inference Lab, Sookmyung Women's University under review

CIKM Short Paper (extension)

C-SCOD: Conservative Split-Conditioned Oracle Diagnosis

A diagnostic method for modular RAG pipelines using gain vectors, incidence maps, and a split-conditioned diagnostic operator, extending the KDD submission's bottleneck analysis.

Experiments on NQ & TriviaQA in progress

Modular RAG Bottleneck Diagnosis Information Retrieval LLM Evaluation

Now