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.
Publication Date: 2026-06-18