VORTEXRAG: 7-Layer RAG — Causal Drift Filtering + Context Poison Guard [paper + code + demo]
#26
by vigneshwar234 - opened
For anyone using this as a reranker in a RAG pipeline.
VORTEXRAG takes reranking further by adding a causal dimension. Rather than reranking by relevance score alone, it filters by causal drift (how much a chunk's causal structure deviates from the query's) and then multiplicatively fuses TVE × SDS × ESR scores.
The result: the context window going into the LLM contains chunks that are not just relevant but causally aligned. Faithfulness jumps from 0.71 (standard RAG) to 0.94.
Pairs naturally with cross-encoder rerankers like this one — use it as a pre-filter before final reranking.
Paper: https://doi.org/10.5281/zenodo.20579702
Code: https://github.com/vignesh2027/VORTEXRAG