Text-to-Solr: A Hybrid Retrieval System for Forensic Investigations

This thesis presents the design, implementation, and evaluation of an Information Retrieval system for digital forensic investigations that integrates a Large Language Model with theApache Solr search engine. The LLM translates natural language investigative queries into two complementary representations: a structured Solr query for lexical, Boolean, and metadata-aware retrieval, and a semantic query for embedding-based similarity search. Keeping these representations separate supports a hybrid retrieval workflow in which explicit constraints and inspectable search logic are combined with semantic matching over dense vector representations. The system is evaluated in an eDiscovery oriented scenario by comparing different LLMs and prompting strategies in terms of valid structured Solr query generation and retrieval effectiveness.