For research and information-retrieval evaluation only. Not a medical device. Not for diagnosis, treatment, or patient care.

RareRAG Bench · v1 release

A research benchmark for RAG over rare-disease literature

RareRAG Bench measures how reliably retrieval and RAG systems find, combine, and cite sparse scientific evidence for rare-disease questions. It is a benchmark, not a product — every question is answered only in terms of measured retrieval/generation behavior on a frozen corpus, never in terms of clinical truth.

Diseases
35
real MONDO/Orphanet IDs
Corpus documents
832
PubMed abstracts, frozen
Benchmark tasks
45
task types A/B/C/D/E/H
Claim support rate
89.6%
n=125 graded claims

What this is — and isn’t

This is
  • A frozen-corpus IR/RAG benchmark: retrieval, reranking, generation, and citation-grading measured against real gold evidence.
  • An honest measurement instrument — every reported difference is bootstrap-tested, and non-significant results are reported as such, not hidden.
  • A first-class abstention benchmark: “insufficient evidence” is a valid, gradeable answer, not a failure mode.
This is not
  • A medical device, diagnostic tool, or patient-care system.
  • A source of personalized diagnosis or treatment advice — no question here asks “what does this patient have.”
  • A statement about clinical validity — results describe IR/RAG behavior on one frozen corpus and task set only.
The headline finding

A null result that motivated its own fix

At v0 scale (20 diseases, 18 gold-bearing tasks), every retrieval comparison this project ran — hybrid vs. dense, hybrid vs. BM25, reranking vs. pre-rerank hybrid — produced descriptively promising differences that did not survive a paired bootstrap significance check. Every 95% CI touched zero. Rather than report those numbers as findings, the project treated this as a scale problem and expanded the benchmark to 35 diseases and 45 tasks specifically to address it (Phase 11).

At v1 scale, two of those comparisons crossed into real, CI-backed significance:

Hybrid beats BM25 on Recall@10
+0.090 [0.007, 0.186]
95% CI excludes zero
Reranking beats pre-rerank hybrid on Recall@10
+0.114 [0.048, 0.190]
95% CI excludes zero

Hybrid vs. dense on Recall@10 (diff -0.036, CI [-0.114, 0.038]) and every nDCG/MRR comparison remain not distinguishable from noise even at n=35 — expanding the task set was not a guarantee of significance across the board, only for two of the tested comparisons. See the full leaderboard for every comparison with its actual CI.

Central research questions

Every question is answerable only in terms of measured retrieval/generation behavior — “the system got the right answer” is not a valid framing; “the system retrieved/cited evidence that a grader judged supportive” is.

RQ1
Which retrieval strategies work best when relevant evidence is sparse?
RQ2
Does hybrid retrieval outperform dense-only retrieval?
RQ3
How much does query expansion improve recall? — not yet run
RQ4
When does reranking help, and what is the first-stage recall ceiling it can't exceed?
RQ5
How reliably do generated answers cite evidence that actually supports them?
RQ6
Can abstention reduce unsupported medical claims?
RQ7
How does performance vary with disease rarity and corpus size?
RQ8
Can structured biomedical knowledge (KG) improve literature retrieval? — optional/stretch, not yet run
Benchmark leaderboard →
BM25 / dense / hybrid, reranking, and every significance test.
Task explorer →
All 45 tasks with abstention and citation outcomes.
Methodology →
Task types, metrics, corpus construction, chunking.