Hybrid RAG for Portuguese Clinical Text

The public Hybrid RAG work treats BM25 and dense retrieval as complementary methods for Portuguese clinical text, fused with Reciprocal Rank Fusion. This note restates only what is already on the site and in the Zenodo record.

The engineering page describes the system as an engineering lab: the architecture, the experiments, and the architecture decision records sit next to the code. The constraint named there is ranking quality versus serving cost. The recorded decision is Reciprocal Rank Fusion over a reranker.

Problem

Clinical retrieval in Portuguese

Constraint

Ranking quality vs. serving cost

Decision

RRF over a reranker

Evidence

Paper · Dataset · Repository

What the paper reports

The companion paper is deposited at DOI 10.5281/zenodo.19686739, titled BM25 and Dense Retrieval Are Complementary for Portuguese Clinical Text: An Empirical Study of Hybrid RAG Across 500 Clinical Queries (Eduardo, 2026; Zenodo).

The research page states the finding in one sentence: BM25 and dense retrieval are complementary, not competing, for Portuguese clinical retrieval. The study is evaluated across 500 clinical queries.

Public artifacts

The paper is paired with a public repository and a Hugging Face dataset, both linked from the homepage. The repository is documented as a companion implementation of the paper: BM25, semantic search, and Reciprocal Rank Fusion. Architecture notes and ADRs are published in the same repository.