Autonomous Knowledge Agent From Query to Report Without Human Intervention

Abstract

This thesis presents a locally executed autonomous literature search tool that treats the user’s research idea as protected information. In this thesis, autonomous means that after the user provides a research query and validates the interpreted requirements, the system continues the literature search process without further human intervention. The system combines four main components. The Agentic Document Searcher sources documents. The Document Ingestion system converts documents into searchable chunks . The Agentic RAG retrieves potentially relevant chunks and documents based on the information requirements derived from the user query. It uses an Information Foraging Theory-inspired classifier to distinguish between answering, interesting, and unrelated chunks, allowing the system to prioritize promising information patches and perform document deep dives when needed. The Knowledge Orchestrator coordinates the workflow from requirement interpretation to retrieval and final report generation. The intuition behind the system is its design to run locally, ensuring stronger control over the user query when doing a literature search. The result is a working prototype that demonstrates how literature search can be automated while maintaining privacy for protected research input. At the same time, the evaluation indicates that performance remains limited by source coverage, relevance assessment, and document processing limitations.