Explainability Analysis of Retrieval-Driven Behavior in RAG Pipelines

This work presents a systematic explainability analysis of retrieval driven behavior in Retrieval Augmented Generation RAG pipelines. The study examines how embedding selection, FAISS based vector retrieval, and generator architectures collectively influence answer correctness, reasoning stability, and hallucination behavior. Using controlled experiments on the SQuAD v2 dataset, the analysis quantifies retrieval precision, semantic drift, and error propagation across the pipeline. Multiple explainability methods are applied, including attention analysis, integrated gradients attribution, and confidence calibration, to trace how retrieved evidence is consumed by the generator. The results show that retrieval precision is the dominant factor governing RAG reliability. When retrieval is semantically aligned, the generator produces stable and grounded outputs. When retrieval drifts or becomes ambiguous, output variance and error rates increase sharply. These findings highlight the importance of retrieval centric evaluation and provide actionable insights for designing more transparent and robust RAG systems.