RESEARCH · RESEARCH · #841
EvidenT: Building trustworthy enterprise assistants through evidence verification (arXiv:2609.22537v1)
EvidenT is a lightweight pipeline that verifies extracted evidence against retrieved documents before answer generation, using structured passage extraction and deterministic lexical alignment to filter unsupported content, correct citation drift, and preserve source-span traceability. On ~500 real enterprise queries the paper reports a 29% average improvement in gold-source hit rate versus prompting baselines, no citations to nonretrieved URLs, and near-saturated answer-to-source lexical coverage — achieved without model retraining.
KEY POINTS
- EvidenT is a lightweight pipeline that verifies extracted evidence against retrieved documents before answer generation, using structured passage extraction and deterministic lexical alignment to filter unsupported content, correct citation drift, and preserve source-span traceability.
- On ~500 real enterprise queries the paper reports a 29% average improvement in gold-source hit rate versus prompting baselines, no citations to nonretrieved URLs, and near-saturated answer-to-source lexical coverage — achieved without model retraining.
- Reducing unsupported claims and citation drift in RAG pipelines improves verifiability and auditability of enterprise assistants, enabling more trustworthy deployments without retraining.
WHY IT MATTERS
Reducing unsupported claims and citation drift in RAG pipelines improves verifiability and auditability of enterprise assistants, enabling more trustworthy deployments without retraining.