EvidenT: Building Trustworthy Enterprise Assistants through Evidence Groundedness and Traceability

Anubha Kabra; Katie Jooyoung Kim; Colin Zhiwei Kou; Helene Sajer; Yimei Fan; Radomir Cisar; Heather Greenhalgh; Gabriel Martinez Vidiri

cs.AISep 18, 2026

EvidenT: Building Trustworthy Enterprise Assistants through Evidence Groundedness and Traceability

Authors: Anubha Kabra, Katie Jooyoung Kim, Colin Zhiwei Kou, Helene Sajer, Yimei Fan, Radomir Cisar, Heather Greenhalgh, Gabriel Martinez Vidiri

Organizations: Bloomberg

Abstract

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 drift, and preserve source-span traceability. On approximately 500 real enterprise queries, EvidenT improves gold-source hit rate by an average of 29% over prompting baselines, produces no citations to nonretrieved urls, and achieves near-saturated answer-to-source lexical coverage.

Figures & tables

Error Type	Description	Root Cause	Impact
Hallucinated Links	Nonexistent citations or urls	Pattern-based generation	Erodes trust
Citation Drift	Cited passage doesn’t support claim	Misaligned grounding	Reduces factual reliability
Limited Traceability	Hard to locate cited text	Buried content, weak anchors	Lowers transparency

Table 1: Key limitations of the pilot enterprise assistant deployment.

Figure 2: A step-by-step visual of our pipeline showing how a query is processed, from document retrieval to answer generation with example content that is entirely fictional and used only for illustration.

Method	SemMax		SemRecall
	M1	M2	M1	M2
Direct Prompting	3.280	4.100	-0.390	1.930
Ground Every Sentence	4.168	4.320	0.439	2.360
EvidenT	5.360	5.990	4.390	5.450

Table 4: Semantic similarity metrics comparing different model sizes.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

	M1	M2
%hallucinated	0.0	0.0
%drift in url	24.19	13.5
%partially hallucinated	16.4	0.6

Appendix

Table 5: N-gram metrics for models M1: LLaMA-3.1-8B and M2: LLaMA-3.3-70B

	M1	M2
%Relevant	42	59
%Partially relevant	58	87

Appendix

Table 6: Human evaluation results of passage relevance for models M1 and M2.

$n$	$t$	Retention Rate ( $\uparrow$ )	% Hallucination (Pre)	% Hallucination (Post) ( $\downarrow$ )	Cov@10 ( $\uparrow$ )
3	0.5	0.96	8.5	4.1	0.996
3	0.7	0.92	8.5	2.3	0.996
3	0.9	0.87	8.5	1.9	0.993
5	0.5	0.93	10.7	0.9	0.998
5	0.7	0.91	10.7	0.6	0.999
5	0.9	0.81	10.7	0.2	0.999

Appendix

Table 7: Ablation results over

n

-gram size (

n

) and overlap threshold (

t

) evaluated on model M2.

Evidence-Grounded Generation

Evidence-grounded generation produces answers and documents whose claims are supported by traceable sources.

See all 148 papers

AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases
May 7, 2026Susheel Suresh, Hazel Mak, Shangpo Chou +2Tool-Augmented Language Model Agents Agentic Search
Microsoft Corporation
SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation
May 5, 2026Jingxi Qiu, Zeyu Han, Cheng HuangRetrieval-Augmented Generation LLM Answer Verification
ZenWeave AI · Georgetown University · Southern Methodist University
ProvenAI: Provenance-Native Traces of Evidence in Generated Answers
Jun 24, 2026Mohammad Faizan, Dalal AlharthiMulti-Hop QA Data Provenance
College of Information Science University of Arizona

May 7, 2026J/K move · Enter open · S save

AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases

Susheel Suresh, Hazel Mak, Shangpo Chou, Fred Kroon+1

Microsoft Corporation

Tool-Augmented Language Model Agents Agentic Search LLM Agent Harnesses Knowledge-Intensive QA Agentic Retrieval Agentic RAG+4

Open paper AskPDF arXiv

We present AgenticRAG, a practical agentic harness for retrieval and analysis over enterprise knowledge bases. Standard RAG pipelines place significant burden of grounding on the search stack, constraining the language model to a fixed candidate set chosen deep in the retrieval process. Our approach reduces this overdependence by layering a lightweight harness on top of existing enterprise search infrastructure, equipping a reasoning LLM with search, find, open, and summarize tools enabling the model to iteratively retrieve information, navigate within documents, and analyze evidence autonomously. On three open benchmarks we observe substantial gains:

49.6\%

recall@1 on BRIGHT (+21.8 pp over the best embedding baseline), 0.96 factuality on WixQA (

+13\%

relative improvement), and

92\%

answer correctness on FinanceBench--within 2 pp of oracle access to true evidence. Ablation studies show that the most significant factor is the shift from single-shot retrieval to agentic tool use (

5.9\times

improvement), while multi-query search and in-document navigation contribute to both quality and efficiency. We present various design choices in our agentic harness that were informed by pre-production deployments. Our results demonstrate its suitability for real-world enterprise production environments.

Abstract

Evidence-Grounded Generation

Explore similar work

AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases

SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation

ProvenAI: Provenance-Native Traces of Evidence in Generated Answers