Hugging Face's logo

vigneshwar234
/
VORTEXRAG-Framework

Question Answering
English
vortexrag
retrieval-augmented-generation
rag
causal-reasoning
hallucination-reduction
semantic-drift
context-window-poisoning
multi-hop-qa
information-retrieval
nlp
Model card Files
xet
 Community
VORTEXRAG-Framework / README.md
vigneshwar234's picture
vigneshwar234
Add comprehensive VORTEXRAG framework model card
4dbb914 verified
preview code
|
raw
history blame contribute delete
6.34 kB
metadata 
license: mit
language:
  - en
tags:
  - retrieval-augmented-generation
  - rag
  - causal-reasoning
  - hallucination-reduction
  - semantic-drift
  - context-window-poisoning
  - multi-hop-qa
  - information-retrieval
  - nlp
  - question-answering
library_name: vortexrag
pipeline_tag: question-answering

VORTEXRAG Framework

Vector Orthogonal Resonance-Tuned EXtraction Retrieval-Augmented Generation

A unified 7-layer RAG framework that simultaneously eliminates Semantic Drift and Context Window Poisoning — the two compounding failure modes that undermine factual grounding in standard RAG systems.

Key Results

Metric VORTEXRAG vs Naive RAG vs CRAG vs Self-RAG
EM 74.8 +13.6 +7.9 +6.4
F1 82.6 +14.2 +8.3 +6.7
Faithfulness 0.94 +0.23 +0.16 +0.13
Semantic Drift Reduction 61% — — —
Context Poison Reduction 71% — — —
Added Latency 45ms — 2.5× faster 2.2× faster

Evaluated on NQ + HotpotQA + MuSiQue + 2WikiMultiHopQA (31,240 total questions).

The 7-Layer Pipeline 

Query
  │
  ▼
[L1: TVE] Tri-Vector Encoding
  │  v = [α·sem(768d); β·syn(64d); γ·cau(32d)]
  │  Encodes text as orthogonal semantic+syntactic+causal vectors
  │
  ▼
[L2: VRC] Vortex Retrieval Cone
  │  spiral_rank = TVE·e^{−λr}·cos(nθ)
  │  Geometric suppression of causally orthogonal chunks (θ > 45°)
  │
  ▼
[L3: SDC] Semantic Drift Corrector      ← per-chunk causal gate
  │  SDS = 1 − tanh(‖v_cau(q) − v_cau(c)‖ / τ) ≥ 0.72
  │  Eliminates individual semantic drift
  │
  ▼
[L4: CPG] Context Poison Guard          ← window-level quality gate
  │  ESR = Σ SDS·w / (P+ε) ≥ 3.5
  │  Greedy-optimal purging (Theorem 5.1)
  │
  ▼
[L5: RFG] Rank Fusion Gate
  │  Φ = TVE^α × SDS^β × ESR_contrib^γ  (multiplicative, no-weak-link)
  │
  ▼
[L6: CCB] Causal Context Builder
  │  pos = rank(Φ+) × causal_depth
  │  Root-cause chunks at pos=0 (U-shaped LLM recall exploitation)
  │
  ▼
[L6: LLM] Generation
  │
  ▼
[L7: FV] Faithfulness Verifier ←──────────────── regeneration loop ──┐
  │  ΔR = 1 − ROUGE-L × NLI ≤ 0.15                                   │
  │  DeBERTa-v3-small CrossEncoder NLI                                 │
  └─── if ΔR > δ_FV: re-weight RFG → retry (max 3 iterations) ────────┘
  │
  ▼
Answer* (argmin ΔR across iterations)

Quick Start 

pip install vortexrag

from vortexrag import VortexRAG, VortexConfig

# Initialize with domain preset
config = VortexConfig(domain="general")  # general, medical, legal, financial, code...
rag = VortexRAG(config)

# Index your documents
rag.index(["Document 1...", "Document 2...", "Document 3..."])

# Query
result = rag.query("Why did X cause Y rather than Z?")
print(result.answer)
print(f"Faithfulness: ΔR={result.delta_r:.3f}")
print(f"Context Quality: ESR={result.esr:.3f}")

Domain Presets

VORTEXRAG ships with 11 pre-calibrated domain parameter vectors:

Domain τ θ_CPG γ (causal) β (syntactic) Use Case
general 0.80 3.5 0.25 0.25 Default balanced
medical 0.35 5.0 0.40 0.15 Drug mechanisms, clinical QA
legal 0.40 4.5 0.35 0.30 Precedent chains, statutory analysis
scientific 0.30 4.0 0.40 0.20 Physics, chemistry, biology
financial 0.50 3.5 0.30 0.25 Market causation, risk analysis
code 0.60 3.5 0.25 0.45 Debugging, AST-structured retrieval
cybersecurity 0.45 4.0 0.35 0.30 Exploit chains, threat intel
educational 0.65 3.0 0.25 0.20 Concept progression, tutoring
historical 0.90 3.0 0.35 0.20 Event causation chains
creative 1.20 2.5 0.15 0.20 Thematic retrieval

Theoretical Contributions

  • Theorem 5.1 (CPG Greedy Optimality): Per-step removal of argmin SDS maximizes ΔESR. Proof via monotone derivative argument.
  • Corollary 5.1 (Convergence): Purge terminates in ≤|W|−3 steps with strictly monotone increasing ESR.
  • Proposition 10.1 (TVE Orthogonality): Cross-arm correlation ρ < 0.08 empirically via Johnson-Lindenstrauss.
  • CCB Positional Optimality: Optimal under U-shaped recall model f(pos) ≈ ½(1+cos(π·pos/L)) (Liu et al. 2023).

Ablation Results

Every layer contributes:

Layer Added EM ΔEM Insight
Baseline 61.2 — Standard cosine RAG
+ TVE 65.3 +4.1 Causal encoding separates mechanism from consequence
+ VRC 67.8 +2.5 Geometric filtering of causally orthogonal docs
+ SDC 70.4 +2.6 Per-chunk SDS gate eliminates individual drift
+ CPG 72.1 +1.7 Window ESR constraint (+39pp context poisoning reduction)
+ RFG 73.4 +1.3 Multiplicative no-weak-link fusion
+ CCB 73.9 +0.5 Root-cause chunks at attention-peak position
+ FV 74.8 +0.9 Faithfulness gate with regeneration loop

Links

Citation 

@article{vignesh2026vortexrag,
  title   = {{VORTEXRAG}: Vector Orthogonal Resonance-Tuned EXtraction
             Retrieval-Augmented Generation},
  author  = {Vignesh L},
  year    = {2026},
  month   = {May},
  doi     = {10.5281/zenodo.20285144},
  url     = {https://github.com/vignesh2027/VORTEXRAG},
  note    = {Independent Research Preprint. v2.0. MIT License.},
  keywords= {RAG, Semantic Drift, Context Window Poisoning, Causal NLP,
             Multi-Hop QA, Faithfulness Verification}
}

Author: Vignesh L — Independent Researcher
ORCID: https://orcid.org/0009-0004-9777-7592
License: MIT
Version: v2.0 — May 2026