TriFuseRAG: Tri-Modal Hybrid Retrieval-Augmented Generation Using SQL, Vector and Graph Databases

Description

Retrieval-Augmented Generation (RAG) systems 
typically route all queries to a single retrieval modality, 
limiting utility for heterogeneous query types that require 
simultaneous structured lookup, entity-relationship 
traversal, and text similarity search. 
We present TriFuseRAG, a prototype multi-modal RAG 
system integrating three retrieval backends—a relational 
SQL engine, an in-memory knowledge graph, and a TF
IDF vector store—under a dual-stage adaptive router and 
weighted fusion layer. A two-tier safety gate filters 
adversarial inputs before retrieval. On a 1,126-query 
closed-world benchmark, the dual-stage router achieves 
98.84% route accuracy. In strict no-leakage evaluation, 
TriFuseRAG achieves 91.3% overall answer accuracy 
(98.1% on supported queries), and 93.9% on composite 
Hybrid queries where all single-source baselines score 
0%. Overall, TriFuseRAG outperforms the best single
source baseline by 34.7 points (91.3% vs. 26.6%), with 
p95 query latency under 10ms on CPU hardware. We 
report full error analysis, latency profiles, and discussion 
of synthetic evaluation scope. 

Authors

DOI: 10.5281/zenodo.20744070

Publication Date: 2026-06-18

Back to publications list


About