Exploring the Potential of Typed Decision Models: Where and How Can Decisions Help in RAG? An Evaluation of Chunk Merging and Candidate Filtering

Evaluations of typed decision models on Brazilian Portuguese remain limited. We evaluate two applications in Portuguese scientific-text retrieval-augmented generation (RAG): filtering retrieved candidates and merging adjacent chunks by idea. The study covers four pipeline arms and 199 questions about 109 of 358 Portuguese SciELO research articles. The corpus representations comprise 17,491 chunks: 10,624 conventional and 6,867 Clef-merged. In the primary full-text experiment, the complete reference passage appeared in the top five for 183/199 questions with unfiltered Clef and 168/199 with the conventional baseline. Filtering changed full-text top-five counts by −6 for the baseline and −24 for Clef. An earlier limited-input experiment, in which a pre-existing 1,200-character candidate limit was identified after the run, is reported separately. Within that experiment’s arms, filtering increased baseline hits by 25 and Clef hits by 37; Clef plus filter reached 144/199 hits. Answer adequacy was assessed separately. The corpus and both chunk collections are provided with this record for further research. Copyright in the source articles and other third-party materials remains with their respective rights holders. No ownership claim is made over those materials. License and attribution details are provided in LICENSES.md and in the corpus inventory.