Work / research

Benchmarking Retrieval for Medical RAG

Head-to-head evaluation of sparse, dense and graph-based retrieval on a PubMed medical corpus, run at scale on WPI's Turing cluster.

Kind
research
Status
ongoing
Role
undergraduate researcher
Period
Jan 2026 – present
Affiliation
WPI, Prof. Gerych's group
domains
nlp information-retrieval medical
methods
rag tf-idf bm25 graph-retrieval dense-retrieval
stack
python pytorch huggingface slurm turing-arc

Problem

A retrieval-augmented generation system is only as good as the documents it retrieves, but “which retriever?” is usually answered by convention rather than by measurement. In the medical domain a bad retrieval is not a bad answer, it is a wrong one. This work asks whether the retrieval methods people default to actually hold up on medical corpora, where terminology is dense, synonymy is rampant and the cost of a miss is high.

Data

  • Corpus: PubMed-derived medical document corpus (Hugging Face)
  • Evaluation: SCT-Bench
  • Splits:
  • Known biases:

Approach

The comparison holds the generator fixed and varies only the retriever, so any difference in downstream answer quality is attributable to retrieval:

  • TF-IDF: sparse lexical baseline
  • BM25: the standard sparse benchmark
  • Graph-based retrieval: exploits structure between documents rather than treating the corpus as a bag of independent passages

Benchmarking jobs are dispatched as Slurm batch arrays on WPI’s Turing ARC cluster, which is what makes a full sweep across retrievers × query sets tractable rather than a weekend of laptop time.

Experiments

What didn’t work

Artifacts