VORTEXRAG: 7-Layer RAG — Causal Drift Filtering + Context Poison Guard [paper + code + demo]
#3
by vigneshwar234 - opened
Relevant for hybrid sparse+dense retrieval research.
VORTEXRAG works as a post-retrieval filtering layer on top of any retrieval backbone — sparse (BM25, SPLADE), dense, or hybrid. The SDC and CPG layers operate on the retrieved candidates, not the retrieval method itself.
This means you can pair your preferred retrieval method with VORTEXRAG's causal filtering to get both lexical precision and causal coherence.
Results with dense retrieval: EM 74.8, Faithfulness 0.94. The gains are largest on causal queries where lexical overlap misleads standard retrieval.
Paper: https://doi.org/10.5281/zenodo.20579702
Code (MIT): https://github.com/vignesh2027/VORTEXRAG