Typed Evidence Fusion for Cross-Modal Enterprise Question Answering: Preserving Provenance, Currency and Cardinality Across Retrieval Modalities

Vishnu Vardhan Reddy Yeruva

doi:10.36948/ijfmr.2026.v08i05.89475

Typed Evidence Fusion for Cross-Modal Enterprise Question Answering: Preserving Provenance, Currency and Cardinality Across Retrieval Modalities

Author(s)	Vishnu Vardhan Reddy Yeruva
Country	United States
Abstract	Enterprises have accumulated a stack of partial solutions to the question-answering problem: semantic search over documents, knowledge graphs over master data, text-to-SQL over the warehouse, and increasingly a chat interface over each of them separately. Each works within its competence and fails outside it, and the failures are silent. Ask a document-RAG system for a total and it will find a sentence containing a number. Ask a text-to-SQL system why a figure moved and it will produce a query, not an explanation. The user cannot see which system they are talking to, so they cannot calibrate. This paper proposes a unified architecture in which retrieval modality is a routed decision rather than a deployment choice, and in which heterogeneous evidence (passages, graph paths, query result sets, computed aggregates) is carried in a typed representation that permits a single verification pass over all of it. The organising abstraction is a shared blackboard over which a planner, four specialist retrieval agents, a critic, and a synthesiser coordinate. We contribute: a typed evidence algebra with composition rules, which makes cross-modal evidence combinable rather than merely concatenable; a coverage-driven termination criterion that decides when to stop retrieving on the basis of what the question requires rather than a fixed step budget; a critic formulation that checks for the specific failure mode of cross-modal inconsistency, where a passage and a query result disagree; and an analysis of the cost–capability curve showing where marginal agent steps stop paying. We evaluate on EntQA-Bench, an executable benchmark released with this paper: a real SQLite warehouse of seven joined tables, 1,332 documents indexed with BM25, a 159-node entity graph, and 1,310 questions in seven classes by required modality, with gold answers computed from the sources. The results are measured. Flattening evidence to text costs accuracy: prose answers 87.6% of questions and a concatenated agent transcript 54.2%, against 100% for typed evidence, and on questions requiring all three modalities the transcript representation answers none. One claim from an earlier simulated draft does not survive. Typed evidence gives no accuracy advantage over plain JSON carrying values and confidences: both answer 100%, and the earlier figure of 8.3 accuracy points is withdrawn. What types buy is verification capability, which the untyped representations lack entirely: source disagreement is detected on 57.2% of injected conflicts against 0.0%, cardinality is reported correctly on 82.5% against 0.0%, and the weakest currency among combined sources is propagated on 100% against 0.0%. The revised claim is therefore narrower and, we think, more useful: types are what allow a cross-modal answer to be checked, not what make it correct.
Keywords	Multi-agent systems, blackboard architecture, knowledge graphs, semantic search, text-to-SQL, enterprise data integration, large language models, evidence fusion, orchestration.
Field	Computer > Artificial Intelligence / Simulation / Virtual Reality
Published In	Volume 8, Issue 5, September-October 2026
Published On	2026-10-11
DOI	https://doi.org/10.36948/ijfmr.2026.v08i05.89475

View / Download PDF File

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

Typed Evidence Fusion for Cross-Modal Enterprise Question Answering: Preserving Provenance, Currency and Cardinality Across Retrieval Modalities

Share this