In enterprise settings, AI assistants must provide answers that can be verified and traced back to original evidence. Conventional retrieval‑augmented generation (RAG) over heterogeneous corporate data often suffers from citation drift, unsupported content, and weak source traceability. EvidenT (T = Trust + Transparency + Traceability) introduces a lightweight pipeline that verifies extracted evidence against the retrieved documents before answer generation, without any model retraining. It first performs structured passage extraction and then applies deterministic lexical alignment to ensure the evidence matches the source text. This step filters out unsupported fragments, corrects citation drift, and preserves span‑level traceability. In experiments on roughly 500 real enterprise queries, EvidenT improves the gold‑source hit rate by about 29%, produces no citations to non‑retrieved URLs, and achieves near‑saturated lexical coverage between answers and sources.
Review