dantunes6/lean-rag-indexes

Lean RAG Indexes for BioASQ

Paper: Retrieval-Bound Generation: Lean RAG Pipelines for Biomedical QA — CLEF 2026 Working Notes, BioASQ Task 14b

Code: github.com/lasigeBioTM/BioASQ14Taskb_2026

Prebuilt retrieval indexes for the Lean RAG Pipelines for Biomedical Question Answering project, developed as part of an MSc dissertation at LASIGE, University of Lisbon (in preparation).

These indexes support a hybrid (BM25 + dense retrieval) pipeline evaluated on BioASQ Task 14b. Everything here was built to run on a single NVIDIA A30 GPU, so the configurations below favour resource efficiency and local deployability over peak retrieval performance.

Repository Structure

faiss/
└── pubmed2026_ivfsq8_raw_matryoshka.index   # Dense vector index
pisa/
├── pubmed2026_soStopw/                      # BM25 index, stopword removal only
│   └── ...
└── pubmed2026_comDois/                      # BM25 index, stopword removal + Porter2 stemming
    └── ...
corpus/
└── pubmed2026.lmdb                          # Key-value store for document text
    └── ...
jsonl2026/
    └── ...                                  # PubMed 2026 Annual Baseline, JSONL format

Corpus

All indexes are built over the PubMed Annual Baseline 2026 (released 30 January 2026), comprising approximately 35 million documents. Only the title and abstract of each record are indexed.

Indexes

FAISS (Dense Retrieval)

File	faiss/pubmed2026_ivfsq8_raw_matryoshka.index
Index type	IVF + Scalar Quantizer (SQ8)
Clusters	216, trained on a random sample of vectors from the corpus
Similarity metric	Inner product
Embedding model	NeuML/pubmedbert-base-embeddings-matryoshka
Embedding dimension	768
Encoded text	title + abstract
Recommended nprobe	512 — tuned on the BioASQ training set for the best trade-off between Recall and per-query latency

PISA — Stopwords only (pisa/pubmed2026_soStopw/)

Retrieval model	BM25
Preprocessing	Stopword removal (PyTerrier / Terrier default stopword list)
Stemming	None
Recommended BM25 parameters	k1=0.6, b=0.4

PISA — Stopwords + Porter2 (pisa/pubmed2026_comDois/)

Retrieval model	BM25
Preprocessing	Stopword removal (PyTerrier / Terrier default stopword list)
Stemming	Porter2
Recommended BM25 parameters	k1=0.6, b=0.4

This is the configuration used for all reported results. A grid search over k1, b and four preprocessing configurations (none, stemming only, stopwords only, both) was run on the BioASQ training set targeting Recall; stopword removal combined with Porter2 stemming came out on top. The soStopw index is included for ablation purposes.

The BM25 parameters above are not PyTerrier's defaults (k1=1.2, b=0.75). Using the defaults will not reproduce the reported results.

Usage

Download

from huggingface_hub import snapshot_download

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

The full set of files is large. To fetch only what you need, use allow_patterns:

snapshot_download(
    repo_id="dantunes6/lean-rag-indexes",
    repo_type="dataset",
    local_dir="./lean-rag-indexes",
    allow_patterns=["faiss/*"]
)

Loading the FAISS index

import faiss

index = faiss.read_index("faiss/pubmed2026_ivfsq8_raw_matryoshka.index")
index.nprobe = 512  # recommended

Queries must be encoded with the same model used to build the index (NeuML/pubmedbert-base-embeddings-matryoshka) and normalised consistently with it, since the index uses inner product as its similarity metric.

Using the PISA index

The PISA index is used through pyterrier_pisa. Point your retriever at the index directory after downloading, and set the BM25 parameters explicitly:

from pyterrier_pisa import PisaIndex

index = PisaIndex("pisa/pubmed2026_comDois")
bm25 = index.bm25(k1=0.6, b=0.4, num_results=1000)

Corpus Notice

The corpus used to build these indexes consists of PubMed abstracts from the PubMed Annual Baseline 2026, distributed by the National Library of Medicine (NLM).

NLM does not endorse or claim any responsibility for derived works, and the data is not covered by copyright in the United States — but records may contain material subject to third-party rights, and NLM's Terms and Conditions apply. Please make sure you comply with them before redistributing this data.

Citation

If you use these indexes, please cite the working notes:

@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}
}

The permanent CEUR-WS volume number and URL will be added once the proceedings are published (September 2026).

Credits

Developed at LASIGE, University of Lisbon, by Diogo Antunes. Supervised by Francisco M. Couto.

PS

I am continuously looking to improve the quality of this dataset. If you spot any inconsistencies, errors, or anomalies in the data, please feel free to report them by opening a new discussion in the Community tab. Your feedback is highly appreciated!