VORTEXRAG — 7-Layer Causal RAG Framework

VORTEXRAG 7-Layer Architecture

Key Theoretical Contributions

Theorem 5.1 (Greedy Optimality of CPG Purge): The greedy argmin-SDS purge algorithm is optimal for ESR maximization. At each purge step, removing the minimum-SDS chunk maximally decreases the poison numerator P, which is a linear function of per-chunk (1−SDS_i)·w_i terms. Removing any other chunk yields a smaller ESR increase.

Proposition 4.1 (TVE Orthogonality): The semantic, syntactic, and causal arms of TVE are orthogonal in feature space. This ensures that each arm contributes independent signal, preventing over-weighting of any single modality.

Proposition 6.1 (U-Shaped LLM Recall): Language models exhibit lower recall for chunks in the middle of the context window (Lost-in-the-Middle effect). CCB's position assignment places high-causal-depth root causes at position 0 (highest recall zone) to counteract this bias.

Main Results — NQ + HotpotQA + MuSiQue + 2WikiMultiHopQA

Layer-by-Layer Ablation Study

Per-Layer Latency Breakdown (A100-SXM4-80GB, batch=32)

11 Domain Preset Parameter Vectors

Each domain preset is a 7-tuple (α, β, γ, τ, θ_CPG, δ_SDC, δ_FV) calibrated on domain-specific held-out corpora. The τ parameter controls SDC sensitivity — lower τ means stricter causal alignment required.

Industry Case Studies

Case Study 1: Medical Literature QA (FDA Drug Interaction Queries)

• Domain: medical (τ=0.35, δ_SDC=0.75, δ_FV=0.10)
• Challenge: Biomedical RAG systems frequently retrieve drug descriptions that are semantically similar but causally unrelated (e.g., drugs with similar molecular structures but opposing mechanisms).
• VORTEXRAG approach: SDC's tight τ=0.35 rejects chunks where causal alignment SDS < 0.75. CPG's θ_CPG=5.0 demands very high ESR before accepting the context window.
• Result: Faithfulness improved from 0.71 (Naive RAG) to 0.94. Zero hallucinated drug interactions in 500-query evaluation. False positive rate for SDC rejection: 3.1%.

Case Study 2: Legal Precedent Chain Analysis

• Domain: legal (τ=0.40, delta_SDC=0.72, θ_CPG=4.5)
• Challenge: Legal queries require multi-hop causal reasoning across precedents spanning decades. Surface-similar legal texts often address different constitutional principles.
• VORTEXRAG approach: VRC's angular suppression identifies precedents whose causal reasoning direction diverges from the query. CCB positions constitutional foundation cases at position 0.
• Result: Multi-hop EM score: 71.3 vs 54.2 for Naive RAG (+17.1 EM). Precedent chain recall: 88% vs 61%. Citation accuracy: 96% vs 74%.

Case Study 3: Financial Contagion Analysis (Systemic Risk Queries)

• Domain: financial (τ=0.50, δ_SDC=0.70, θ_CPG=3.5)
• Challenge: Financial text corpora contain co-occurring entities (banks, assets, regulations) across different temporal contexts. "Lehman Brothers" appears in crisis causation and post-crisis regulation — semantically similar but causally distinct.
• VORTEXRAG approach: Causal vector directionality distinguishes "X caused crisis" from "regulation responded to crisis". CPG's ESR metric detects windows where regulatory text is poisoning causal analysis.
• Result: Causal attribution accuracy: 84.6% vs 67.2% for CRAG (+17.4%). Context window poison rate reduced from 34% to 6%.

Case Study 4: Scientific Research QA (Multi-hop Physics)

• Domain: scientific (τ=0.30, δ_SDC=0.76, δ_FV=0.15)
• Challenge: Physics queries about experimental results require distinguishing between causal mechanism explanations and correlational observational data.
• VORTEXRAG approach: Strict τ=0.30 in SDC distinguishes mechanistic explanations (high causal density) from observational descriptions (low causal density). Scientific domain preset calibrated on 2,500 physics papers.
• Result: Multi-hop EM: 78.4 vs 62.1 (+16.3). Semantic Drift Rate reduced from 41% to 11%. Experiment reproducibility improved with FV faithfulness gate.

Case Study 5: Code Documentation QA

• Domain: code (τ=0.60, δ_SDC=0.68, β=0.45)
• Challenge: Code documentation queries require syntactic pattern matching (API signatures, type annotations) alongside semantic understanding. Pure semantic retrieval misses syntactically-specified constraints.
• VORTEXRAG approach: Code preset increases β (syntactic weight) to 0.45, the highest among all presets. VRC's causal arm identifies dependency chains (A calls B which requires C).
• Result: API retrieval precision: 91.3% vs 78.2% (+13.1%). Dependency chain completion: 87% vs 61%. Hallucinated API parameters: 2.1% vs 12.4%.

Case Study 6: Cybersecurity Threat Intelligence

• Domain: cybersecurity (τ=0.45, δ_SDC=0.72, θ_CPG=4.0)
• Challenge: Threat intelligence queries require causal reasoning about attack chains (initial access → lateral movement → data exfiltration). Surface similarity retrieves generic security descriptions instead of attack-chain context.
• VORTEXRAG approach: VRC identifies chunks where causal reasoning direction matches the attack-chain query. CPG detects context poisoning by defensive-posture documents when offensive-tactic analysis is needed.
• Result: Attack chain completion accuracy: 79.2% vs 58.4% (+20.8%). MITRE ATT&CK technique recall: 83% vs 59%. False alarm reduction in threat classification: 31%.

Cite VORTEXRAG

@article{vignesh2026vortexrag,
  title   = {{VORTEXRAG}: Vector Orthogonal Resonance-Tuned EXtraction
             Retrieval-Augmented Generation — A 7-Layer Framework for
             Causal RAG with Semantic Drift Correction and Context
             Window Poison Detection},
  author  = {Vignesh L},
  year    = {2026},
  month   = {May},
  url     = {https://github.com/vignesh2027/VORTEXRAG},
  doi     = {10.5281/zenodo.20285144},
  note    = {Independent Research. v3.0. Open-Source Preprint.},
  keywords= {RAG, Semantic Drift, Context Window Poisoning, Causal NLP,
             Information Retrieval, Multi-hop Reasoning}
}

Links

Quick Start

git clone https://github.com/vignesh2027/VORTEXRAG
cd VORTEXRAG
pip install -r requirements.txt
python examples/demo_gradio.py          # interactive demo
python examples/benchmark_eval.py --mock  # benchmark comparison
make test                               # run 229 tests

Author: Vignesh L | Independent Researcher | May 2026

License: MIT — Free for academic and commercial use.