SOURCE-LINKED INTELLIGENCE

EvidenT: Building Trustworthy Enterprise Assistants through Evidence Groundedness and Traceability

arXiv · AI, language, vision and robotics · article · Sep 18, 2026 · UTC

Enterprise AI assistants must produce responses that are verifiable and traceable to source evidence. However, retrieval augmented generation (RAG) over heterogeneous enterprise data can suffer from citation drift, unsupported content, and weak source traceability. We present EvidenT (T = Trust + Transparency + Traceability), a lightweight pipeline that verifies extracted evidence against retrieved documents before answer generation, without model retraining. EvidenT combines structured passage extraction with deterministic lexical alignment to filter unsupported content, correct citation drif

Read original source ↗ Open in workspace

recordType

paper

region

Global