VORTEXRAG: 7-Layer RAG — Causal Drift Filtering + Context Poison Guard [paper + code + demo]
#7
by vigneshwar234 - opened
Sharing recent work that might be relevant to this community.
VORTEXRAG (2026) is a 7-layer RAG framework addressing two problems that persist across all existing RAG systems: Semantic Drift (retrieval of causally-wrong but topically-similar chunks) and Context Window Poisoning (low-signal chunks diluting LLM attention).
Key technical contributions:
- Tri-Vector Encoding: 864-dim = semantic (768d) + syntactic (64d) + causal PropBank (32d)
- Vortex Retrieval Cone: spiral_rank = TVE·e^(−λr)·cos(nθ) for geometric angular suppression
- Context Poison Guard: greedy ESR purge with proven optimality (Theorem 5.1)
Results: EM 74.8, F1 82.6, Faithfulness 0.94 across NQ/TriviaQA/WebQ/PopQA/HotpotQA/2WikiMH.
Preprint: https://doi.org/10.5281/zenodo.20579702
Code: https://github.com/vignesh2027/VORTEXRAG