Self-Contamination in Continuously Updated RAG Systems

This project investigates self-contamination in continuously updated Retrieval-Augmented Generation (RAG) systems: the phenomenon where a RAG system's own previously generated outputs are re-ingested into its knowledge base over time, progressively degrading retrieval quality and accelerating hallucination. Grounded in production-scale infrastructure (Qdrant vector store, Ollama-served embeddings, and the bge-m3 embedding model), the work formalizes the contamination dynamics with formal math and benchmarks retrieval degradation across successive update cycles. It builds on and extends recent literature on model collapse, data-poisoning attacks on RAG (e.

g. , PoisonedRAG), information-cascade effects (Spiral of Silence), and hallucination detection, aiming to propose practical mitigation strategies for production RAG deployments.