VORTEXRAG: 7-Layer RAG — Eliminates Semantic Drift & Context Poisoning (EM 74.8, +13.6 vs Naive RAG)
#27
by vigneshwar234 - opened
Hi all! Sharing VORTEXRAG — a new 7-layer RAG framework specifically relevant to this model.
The problem: Cosine similarity retrieval cannot distinguish causally relevant chunks from topically similar ones. The result: semantic drift (wrong chunks in context) and context window poisoning (irrelevant chunks drowning the right one).
VORTEXRAG encodes text as a 864-dimensional tri-vector (semantic + syntactic + causal), then runs a 7-layer pipeline:
- SDC filters chunks with SDS = 1−tanh(‖D‖/τ) ≥ 0.72 (causal drift gate)
- CPG purges context until ESR ≥ 3.5 (provably optimal greedy algorithm)
- FV post-generation faithfulness check: ΔR = 1−ROUGE-L×NLI ≤ 0.15
Results vs baselines:
| System | EM | F1 | Faithfulness |
|---|---|---|---|
| VORTEXRAG | 74.8 | 82.6 | 0.94 |
| Self-RAG | 68.4 | 77.1 | 0.81 |
| Naive RAG | 61.2 | 69.4 | 0.71 |
If you use this model in a RAG pipeline, VORTEXRAG can dramatically reduce hallucinations.
📄 Paper: https://doi.org/10.5281/zenodo.20579702
💻 Code (229 tests, MIT): https://github.com/vignesh2027/VORTEXRAG
🚀 Demo: https://huggingface.co/spaces/vigneshwar234/VORTEXRAG
vigneshwar234 changed discussion status to closed