Agentic GraphRAG beats vector RAG on 7M Swiss financial docs

A new paper from researchers Arthur Capozzi and Dirk Helbing presents Agentic GraphRAG, a collaborative AI framework designed to navigate complex unstructured financial data. The system first builds a Neo4j knowledge graph via a three-phase pipeline: deterministic ingestion of 'strong nodes' from verified fields, LLM-based extraction of weak nodes from unstructured legal notices, and deterministic identity resolution and deduplication. On top of this graph, a modular analytical agent employs zero-shot intent routing, a bounded reflection loop, secure tool-mediated graph access, and state-aware response synthesis. A human-in-the-loop dashboard provides transparency by exposing evidence and execution traces.

The framework was rigorously evaluated on the Swiss Official Gazette of Commerce, a multilingual corpus of over seven million publications spanning seven years. A multi-tier evaluation protocol measured entity-resolution precision, tool-routing behavior, answer quality, and multi-turn conversational performance. Across automated, human-curated, and conversational benchmarks, Agentic GraphRAG consistently outperformed a standard agentic vector-RAG baseline, achieving strong gains in correctness, answer relevance, information recall, turn success rate, and context carryover accuracy. The architecture is modular, reproducible, and transferable to other commercial gazettes and public-sector registry systems, making it a significant advance for financial data analysis.