VORTEXRAG: 7-Layer RAG β Causal Drift Filtering + Context Poison Guard [paper + code + demo]
Relevant for anyone using this model as a retrieval backbone.
One limitation of all embedding-based retrieval (including this model) is that cosine similarity can't separate causal relevance from topical association. VORTEXRAG addresses this by adding a causal filtering layer on top of embedding retrieval.
Architecture: your embedding model handles the ANN search β VORTEXRAG's SDC/CPG layers filter by causal drift β FV layer verifies faithfulness post-generation.
Combined results with this retrieval approach: EM 74.8, Faithfulness 0.94 (+0.23 over standard embedding retrieval baseline).
The 11 domain presets (medical Ο=0.35, legal Ο=0.40, scientific Ο=0.30, etc.) let you tune the causal strictness per use case.
Paper: https://doi.org/10.5281/zenodo.20579702
Code (MIT, 229 tests): https://github.com/vignesh2027/VORTEXRAG
Demo: https://huggingface.co/spaces/vigneshwar234/VORTEXRAG