Intent-Aware Retrieval-Augmented Generation for EU Legislation: A Multi-Source Framework for Summarizing Laws, Debates, and News

Abstract

Retrieval-augmented generation (RAG) pipelines typically employ uniform retrieval strategies across queries, disregarding document-type imbalance in heterogeneous corpora. This thesis investigates whether intent-aware agentic retrieval with source-type weighting improves query-guided summarization over a multi-source EU legislative corpus compared to a standard RAG pipeline. The system classifies each query into one of four intent types (factual, temporal, stance, comparative) and uses this classification to condition source-type retrieval weighting, query decomposition, and summarization prompt selection. These mechanisms are combined with hybrid BM25-dense retrieval using reciprocal rank fusion. A six-level ablation study over 40 benchmark queries evaluates each retrieval component incrementally. The full system achieves attribution of 0.642, outperforming the standard RAG baseline by 15.5%, with modest trade-offs in alignment (−0.025) and key term coverage (−0.028). Hybrid retrieval achieves perfect alignment (1.000) across all 40 queries, making it the strongest single component. The agentic loop increases attribution further (+0.054) but introduces alignment costs on query types where single-pass retrieval is sufficient. An unexpected interaction between hierarchical chunking and inverse-frequency weighting inverts the document-level class distribution at the chunk level. Human evaluation with three raters confirms that automated metrics capture structural summary properties but not substantive adequacy. The findings demonstrate that hybrid retrieval should be the default for multi-source legislative corpora and that component interactions are intent dependent.