Architecture

Retrieval Pipeline

NFCorpus → validation → lexical + dense retrieval → RRF fusion → cross-encoder reranking → evaluation → statistical analysis → research portal.

Full pipeline

NFCorpus (BeIR/nfcorpus + BeIR/nfcorpus-qrels)

Data Validation

Lexical IR — TF-IDF / BM25

Dense IR — BGE / MedCPT

RRF Fusion (BM25 + MedCPT, rank-based)

MedCPT Cross-Encoder Reranking

Ranked Documents

Qrels (test split)

→

Evaluation (pytrec_eval)

MAP / MRR / nDCG / Precision / Recall

Statistical Analysis (paired bootstrap)

Research Results (this portal)

Stage → module → script

#	Stage	Module	Script
1	Dataset ingestion	biomedical_ir/data.py	scripts/download_data.py
2	Dataset validation	biomedical_ir/validation.py	scripts/audit_dataset.py
3	TF-IDF (M1)	biomedical_ir/tfidf.py	scripts/run_tfidf.py
4	BM25 (M2)	biomedical_ir/bm25.py	scripts/run_bm25.py
5	General dense retrieval (M3, BGE)	biomedical_ir/dense.py	scripts/run_bge.py
6	Biomedical dense retrieval (M4, MedCPT)	biomedical_ir/medcpt.py	scripts/run_medcpt.py
7	Hybrid RRF (M5)	biomedical_ir/fusion.py	scripts/run_hybrid.py
8	Cross-encoder reranking (M6)	biomedical_ir/reranker.py	scripts/run_reranker.py
9	Evaluation + statistics	biomedical_ir/{evaluation,statistics,efficiency}.py	scripts/evaluate_all.py
10	Error analysis	biomedical_ir/error_analysis.py	scripts/error_analysis.py
11	Figures / tables	—	scripts/generate_figures.py
12	Web export	—	scripts/export_web_results.py

Why RRF operates on rank, not raw score

BM25 scores are unbounded and MedCPT similarities are raw (unnormalized) dot products — not on a comparable scale. Summing them directly would implicitly and arbitrarily weight whichever retriever happens to produce larger-magnitude scores, not whichever is more accurate. Reciprocal Rank Fusion instead combines rank positions, which are already on a common scale by construction.

Why Vercel never runs the models

BGE, MedCPT, and the MedCPT cross-encoder are never invoked inside a Vercel serverless function. All embeddings/rankings are computed offline (locally or in Colab), evaluated, and exported as compact JSON artifacts. This portal only reads those artifacts — the /search page is explicitly labeled as showing precomputed experiment output, never a live model call.