README.md

---
title: VORTEXRAG
emoji: "🌀"
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: "5.29.0"
app_file: app.py
pinned: true
license: mit
short_description: "7-Layer RAG: +13.6 EM, 0.94 Faithfulness"
tags:
  - retrieval-augmented-generation
  - RAG
  - NLP
  - question-answering
  - causal-reasoning
  - hallucination-reduction
  - LLM
  - machine-learning
---

# VORTEXRAG

**Vector Orthogonal Resonance-Tuned EXtraction Retrieval-Augmented Generation**

> A 7-layer RAG framework that simultaneously eliminates **Semantic Drift** and **Context Window Poisoning**.

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.20579702.svg)](https://doi.org/10.5281/zenodo.20579702)
[![GitHub](https://img.shields.io/badge/GitHub-vignesh2027%2FVORTEXRAG-blue)](https://github.com/vignesh2027/VORTEXRAG)
[![Tests](https://img.shields.io/badge/Tests-229%20passing-brightgreen)](https://github.com/vignesh2027/VORTEXRAG)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://github.com/vignesh2027/VORTEXRAG/blob/main/LICENSE)

---

## The Problem Standard RAG Cannot Solve

Ask "Why did Lehman Brothers collapse?"

Standard RAG retrieves both Dodd-Frank provisions (cosine 0.87, topically related but WRONG) and the CDS mispricing mechanism (cosine 0.91, causally correct). The LLM sees both and hallucinates a policy-response narrative. **This is Semantic Drift.**

Even with the right chunk retrieved, 7 surrounding irrelevant chunks dilute the LLM attention. **This is Context Window Poisoning.**

VORTEXRAG solves both with a principled 7-layer pipeline.

---

## The 7 Layers

| Layer | Name | Formula |
|-------|------|---------|
| 1 | TVE - Tri-Vector Encoding | score = alpha * cos_sem + beta * cos_syn + gamma * cos_cau |
| 2 | VRC - Vortex Retrieval Cone | spiral = TVE * exp(-lambda*r) * cos(n*theta) |
| 3 | SDC - Semantic Drift Corrector | SDS = 1 - tanh(norm(D)/tau) >= 0.72 |
| 4 | CPG - Context Poison Guard | ESR = sum(S*w)/(P+eps) >= 3.5 (provably optimal) |
| 5 | RFG - Rank Fusion Gate | Phi = TVE^alpha * SDS^beta * ESR^gamma |
| 6 | CCB - Causal Context Builder | pos = rank(Phi) * causal_depth |
| 7 | FV - Faithfulness Verifier | Delta_R = 1 - ROUGE-L * NLI <= 0.15 |

---

## Benchmark Results (v3.0)

| System | EM | F1 | Faithfulness | Latency |
|--------|----|----|-------------|---------|
| **VORTEXRAG** | **74.8** | **82.6** | **0.94** | **185ms** |
| Self-RAG | 68.4 | 77.1 | 0.81 | 410ms |
| CRAG | 66.9 | 75.8 | 0.79 | 320ms |
| Naive RAG | 61.2 | 69.4 | 0.71 | 95ms |

**+13.6 EM** over Naive RAG - Semantic Drift **-61%** - Context Poisoning **-71%** - **2.2x faster** than Self-RAG

---

## Ablation

```
Baseline:          61.2 EM / 0.68 Faithfulness
+TVE:              65.3 EM  (+4.1)
+VRC:              67.8 EM  (+2.5)
+SDC:              70.4 EM  (+2.6)
+CPG:              72.1 EM  (+1.7)
+RFG:              73.4 EM  (+1.3)
+CCB:              73.9 EM  (+0.5)
VORTEXRAG (full):  74.8 EM  (+0.9)
```

---

## 11 Domain Presets

scientific (tau=0.30) - medical (tau=0.35) - legal (tau=0.40) - cybersecurity (tau=0.45) - financial (tau=0.50) - code (tau=0.60) - educational (tau=0.65) - general (tau=0.80) - historical (tau=0.90) - customer (tau=0.95) - creative (tau=1.20)

---

## Citation

```bibtex
@article{vignesh2026vortexrag,
  title   = {VORTEXRAG: Vector Orthogonal Resonance-Tuned EXtraction RAG},
  author  = {Vignesh, L},
  year    = {2026},
  doi     = {10.5281/zenodo.20579702},
  url     = {https://doi.org/10.5281/zenodo.20579702}
}
```

**Paper:** https://doi.org/10.5281/zenodo.20579702
**Code:** https://github.com/vignesh2027/VORTEXRAG
**ORCID:** https://orcid.org/0009-0004-9777-7592