BioASQ Task 14b — Biomedical Question Answering

Participation in the BioASQ Challenge Task 14b, covering both Phase A (document and snippet retrieval) and Phase B (exact and ideal answer generation).

Developed by Diogo Antunes under the supervision of Francisco M. Couto (LASIGE, Faculty of Sciences, University of Lisbon).

For the Working Notes click here

---

Overview

The system is split into two phases:

• Phase A retrieves relevant PubMed documents and extracts answer-bearing snippets. Four interchangeable pipelines were evaluated (see below).
• Phase B takes the Phase A output and generates the exact answer (yes/no, factoid, list) and the ideal answer (summary) with a few-shot prompted LLM.

Phase A pipelines

Pipeline	File	Approach
Hybrid	pipelines/bioasq_pipeline_normal.py	FAISS dense (PubMedBERT) + PISA BM25, weighted-sum fusion, cross-encoder reranking, score-threshold cutoff
Sparse-Only	pipelines/bioasq_pipeline_onlyBM25.py	PISA BM25 only → cross-encoder reranking
DPRF	pipelines/bioasq_pipeline_dprf.py	Dense pseudo-relevance feedback expansion, RRF fusion of BM25 + dense runs, cross-encoder reranking
Ensemble	pipelines/Ensemble_crossencoders.py	Reciprocal-rank fusion across multiple cross-encoders

Phase B

File	Approach
pipelines/phaseb.py	Few-shot prompting with google/gemma-4-E4B-it; two-pass generation (exact answer → ideal answer)

---

Repository structure

.
├── bioasq/                       # shared modules imported by the pipelines
│   ├── helpers.py                
│   ├── corpus_store.py           
│   ├── snippet_extractor.py      
│   ├── sparse_retriever.py       
│   ├── thresholds.py             
│   └── wsum_fuser.py             
├── pipelines/                    # runnable entry points (one per system)
│   ├── bioasq_pipeline_normal.py
│   ├── bioasq_pipeline_onlyBM25.py
│   ├── bioasq_pipeline_dprf.py
│   ├── Ensemble_crossencoders.py
│   └── phaseb.py
├── requirements.txt
└── README.md

---

Setup

# 1. create a virtual environment
python -m venv .venv && source .venv/bin/activate

# 2. install PyTorch matching your CUDA version FIRST
#    see https://pytorch.org/get-started/locally/

# 3. install the rest
pip install -r requirements.txt

Prerequisites (not included in this repo)

Prebuilt indexes (FAISS + PISA) used in this work are available on Hugging Face:

🤗 dantunes6/lean-rag-indexes

Download them with:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="dantunes6/lean-rag-indexes",
    repo_type="dataset",
    local_dir="./lean-rag-indexes"
)

---

Usage

Phase A — example (Hybrid pipeline)

python pipelines/bioasq_pipeline_normal.py \
    --input         path/to/BioASQ-task14bPhaseA-testset.json \
    --output        path/to/submission.json \
    --faiss-index   path/to/pubmed.index \
    --corpus        path/to/pubmed.lmdb \
    --pisa-index    path/to/pubmed_pisa \
    --top-k-retrieval 1000 \
    --top-k-docs 10 \
    --top-snippets 10 \
    --ce-threshold 0.91

Each pipeline exposes --help for its full set of arguments.

Phase B — answer generation

python pipelines/phaseb.py

For the phaseb.py script the paths should be added in the file.

---

Citation

If you use this code, please cite the working notes:

Antunes, D. M. C., Lopes, P. R. C., & Couto, F. M. (2026). Retrieval-Bound Generation: Lean RAG Pipelines for Biomedical QA. In Working Notes of the Conference and Labs of the Evaluation Forum (CLEF 2026), Jena, Germany, pp. 123–145. CEUR-WS.org.

BibTeX

@inproceedings{antunes2026retrievalbound,
  author    = {Antunes, Diogo M. C. and Lopes, Paulo R. C. and Couto, Francisco M.},
  title     = {Retrieval-Bound Generation: Lean {RAG} Pipelines for Biomedical {QA}},
  booktitle = {Working Notes of the Conference and Labs of the Evaluation Forum ({CLEF} 2026)},
  series    = {CEUR Workshop Proceedings},
  publisher = {CEUR-WS.org},
  address   = {Jena, Germany},
  pages     = {123--145},
  year      = {2026}
}

Developed at LASIGE, Faculty of Sciences, University of Lisbon, by Diogo Antunes, supervised by Francisco M. Couto.