Auravector-DB-SynaptiX

AuraVector DB: High-precision, low-latency hybrid vector engine for RAG systems. Combines BGE dense embeddings, Rank-BM25, Reciprocal Rank Fusion (RRF), Cross-Encoder reranking, and SVD entropy analysis to eliminate LLM hallucinations under a p95 < 300 ms constraint. Built for ADROSONIC Build Challenge by Team SynaptiX (BIT Mesra).

Retrieval-Augmented Generation (RAG) connects Large Language Models to private enterprise corpora. However, standard dense-only vector search frequently fails at scale due to semantic vs. factual divergence: dense embeddings retrieve passages that match the general topic but miss exact model codes, product identifiers, or numeric values. Passing these close-but-wrong contexts into an LLM triggers a hallucination cascade, undermining enterprise reliability.