dantunes6/lean-rag-indexes

Lean RAG Indexes for BioASQ

Prebuilt retrieval indexes for the Lean RAG Pipelines for Biomedical Question Answering project, developed as part of a Master's dissertation at LASIGE, University of Lisbon.

These indexes support a hybrid (BM25 + dense retrieval) pipeline evaluated on BioASQ Task 14b.

Repository Structure

faiss/

└── pubmed2026_ivfsq8_raw_matryoshka.index  #Dense vector index

pisa/

├── pubmed2026_soStopw/  #BM25 index with stopword removal only
│   └── ...
└── pubmed2026_comDois/  #BM25 index with stopword removal + Porter2 stemming
    └── ...

corpus/
└──pubmed2026.lmdb  #Embedded database for key-value data
    └── ...

jsonl2026/
    └── ...  #PubMed2026 Annual Baseline

Indexes

FAISS (Dense Retrieval)

• File: faiss/pubmed2026_ivfsq8_raw_matryoshka.index
• Index type: IVF + Scalar Quantizer (SQ8)
• Embedding model: NeuML/pubmedbert-base-embeddings-matryoshka
• Corpus: PubMed Annual Baseline 2026

PISA — Stopwords only (pisa/pubmed2026_soStopw/)

• Retrieval model: BM25
• Preprocessing: Stopword removal using PyTerrier's default stopword list
• Stemming: None
• Corpus: PubMed Annual Baseline 2026

PISA — Stopwords + Porter2 (pisa/pubmed2026_comDois/)

• Retrieval model: BM25
• Preprocessing: Stopword removal using PyTerrier's default stopword list
• Stemming: Porter2
• Corpus: PubMed Annual Baseline 2026

Usage

Loading the FAISS index

import faiss

index = faiss.read_index("faiss/pubmed2026_ivfsq8_raw_matryoshka.index")

index.nprobe= 512 #recommended

Or download it first with:

from huggingface_hub import snapshot_download

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

Using the PISA index

The PISA index is used via PISA. Point your retriever at the pisa/ directory after downloading.

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). Please ensure you comply with NLM's terms of use before redistributing this data.

Credits

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