THE ASSESSMENT OF IMPACT OF PRE-FILTERING ON RETRIEVAL QUALITY IN RAG SYSTEMS WITH VECTOR SEARCH

Abstract

The paper analyzes modern approaches to evaluating Retrieval-Augmented Generation (RAG) systems that integrate vector search with answer generation by large language models (LLMs). It examines classical retrieval quality metrics alongside LLM-oriented generation quality metrics, including their application within frameworks such as RAGAS, ARES, VERA, and MIRAGE. A computational experiment was conducted using a Google Cloud Platform (GCP) Firestore collection with vector search over a dataset of IT professionals' CVs, comparing standard vector search against search enhanced by pre-filtering on metadata. The results demonstrate that pre-filtering increases the proportion of relevant documents in the context, reduces retrieval latency, and enables larger context sizes without proportional degradation in generation quality. The experimental findings confirm the dependence of RAG system answer quality on the purity and relevance of the retrieved context.