EvdenT: Trust and Traceability in Corporate AI Assistants

The responses that AI assistants used in corporate environments should be not only correct, but also traced up to**slip. However, Retrieval Augmented Generation (RAG) systems running on heterogeneous databases often experience slip, unsupported content and poor traceability issues in source display. A new study published in ArXiv offers a light system called EvidenT to bring solutions to these problems.

The main features of EvidenT are:

• Validate evidence from received documents before reply is created
• Eliminate the problems of Citation drift (reflection of source display) and unsupported content (not supported content)
• A lightweight structure that works without requiring fine adjustment of model weights
• Provide full traceability and transparency for outputs

The system is integrated in RAG architecture, widely used in corporate environments. The proven verification step controls that every claim is really located in the sources received before the model produces response. In this way, nominations can be verified and the source for each sentence becomes available.

The light design of EvidenT means it can be easily integrated without affecting existing distributions. In particular, the importance of such solutions is increasing in the sectors of regulation such as law, health and finance. Institutions have to explain and show decisions of AI assistants. Systems like EvidenT help meet this requirement.