Answers are grounded in this paper and cite their sources.
Territorial well-being analysis requires integrating two complementary families of information: objective indicators (equipment, composite quality-of-life scores) that characterize structural conditions, and subjective perceptions (citizen survey verbatims, interviews) that reflect lived experience. Classical retrieval-augmented generation (RAG) handles such mixed-source queries poorly: semantic similarity search fragments information and provides no structural constraint on the consistency between the data types a query calls for and those actually available. We propose a hybrid RAG architecture built on three principles: (i) explicit decomposition of the user query into independently processed sub-questions; (ii) multi-level synthetic representations precomputed along business-driven analytical views, adapting the RAPTOR approach to survey metadata rather than automatic clustering; and (iii) a typed decomposition mechanism enforcing consistency between the nature of the query and the sub-questions generated, constraining the formulation of sub-questions that have no documentary support. We evaluate the architecture on a benchmark of 109 manually designed questions spanning seven usage categories, combining categorical retrieval metrics with a multi-dimensional LLM-as-judge validated against human annotations, an ablation study isolating each principle, and a comparison with an external decomposition baseline (Self-Ask). Results show that the architecture retrieves relevant sources far more exhaustively than vanilla RAG (Recall 0.97 against 0.68), that both the typing rule and the multi-level representations contribute significantly to answer quality, and that it outperforms Self-Ask. On the benchmark as a whole, answer quality is comparable to vanilla RAG, the advantage being concentrated on questions requiring multi-aspect reasoning—which suggests that architectural choices should be informed by the intended usage profile.
No takes yet. Share an insight, caveat, or question.
Comiti et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: