Design of a Flow-Controlled RAG Pipeline for Impact Investment Fund Due Diligence

Summary

Master's thesis (Industrial Engineering, IT major, Haute École ICHEC-ECAM-ISFSC). The work designs and evaluates a flow-controlled Retrieval-Augmented Generation (RAG) pipeline to automate due diligence for impact investment funds, under SFDR, EU Taxonomy and GDPR constraints. Design Science Research methodology, adaptive selection driven by a cumulative-coverage threshold, cross-encoder reranking, and a Privacy-by-Design anonymisation layer. Results: 90.5% accuracy on structured textual documentation and a 90% reduction in preliminary analysis time, while keeping a Human-in-the-Loop. The full text of the thesis is available in French only.

Keywords