VORTEXRAG: 7-Layer RAG — Causal Drift Filtering + Context Poison Guard [paper + code + demo]
#4
by vigneshwar234 - opened
Sharing work that combines knowledge-intensive retrieval with causal reasoning.
VORTEXRAG is a 7-layer RAG framework that encodes causal structure directly into retrieval. The Tri-Vector Encoding (TVE) layer uses PropBank SRL event structure to build a 32-dim causal arm alongside the standard 768-dim semantic arm.
For knowledge-intensive tasks (entity reasoning, multi-hop QA, causal chains), this separation of semantic vs causal retrieval makes a measurable difference: +6.8 EM on HotpotQA, +7.5 EM on 2WikiMultiHopQA vs Self-RAG.
Paper: https://doi.org/10.5281/zenodo.20579702
Code: https://github.com/vignesh2027/VORTEXRAG
Demo: https://huggingface.co/spaces/vigneshwar234/VORTEXRAG