VORTEXRAG: 7-Layer RAG — Causal Drift Filtering + Context Poison Guard [paper + code + demo]
#11
by vigneshwar234 - opened
Relevant for anyone researching hallucination and faithfulness in RAG.
VORTEXRAG attacks hallucination at two stages — retrieval and generation:
- Retrieval stage: SDC layer scores causal drift (SDS = 1−tanh(‖D‖/τ)) and rejects chunks drifting from query causality. CPG layer purges context until Effective Signal Ratio ≥ 3.5.
- Generation stage: FV layer computes ΔR = 1−ROUGE-L×NLI ≤ 0.15. If violated, it reranks context and retries up to 3×.
Faithfulness results: 0.94 (vs Self-RAG 0.81, Naive RAG 0.71). Semantic Drift Rate: 14% (down from 36%). Context Poisoning Rate: 7% (down from 24%).
The FV layer uses DeBERTa-v3-small CrossEncoder for NLI — same family as models on this repo.
Paper: https://doi.org/10.5281/zenodo.20579702
Code (MIT, 229 tests): https://github.com/vignesh2027/VORTEXRAG
Demo: https://huggingface.co/spaces/vigneshwar234/VORTEXRAG