Tax fraud rarely shows up in one document. A payment that passes through two companies before reaching a board member’s relative looks ordinary on every individual invoice. The pattern only appears when you connect the companies, people, accounts, and transactions across hundreds of files.

Redfield built a prototype that does that connecting automatically. LLMs extract entities and relationships from unstructured documents and store them as a graph in Memgraph. Investigators can then search the whole network for suspicious payment patterns and conflicts of interest, or ask questions in plain language, without knowing anything about graph technology. Because the data is sensitive, everything runs locally, from the LLMs to the database.

If you’re building RAG over documents that can’t leave your environment, this session is for you.

Speaker

Jacob Coles is a Data Scientist at Redfield AB. He graduated in mechatronics engineering in Australia, then earned a master’s degree in Language Technology at the University of Gothenburg in Sweden. At Redfield, he helps develop APIs and applications that use machine learning and LLMs to build different types of retrieval-augmented generation (RAG) applications that fuse disparate data sources.

From Documents to Networks: Graph-Based RAG for Tax Investigation