Adaptive Cluster-Graph (ACG) for Efficient Document Retrieval in RAG Pipelines

Traditional Retrieval-Augmented Generation (RAG) systems depend heavily on document retrieval components that need quick retrieval and accurate results when working in knowledge synthesis-intensive fields such as science-based question answering. The traditional retrieval system based on TF and TF-IDF creates simple, understandable retrieval pipelines, even though these methods struggle to find the best semantic associations in information-dense resources. An Adaptive Cluster Graph (ACG) search framework serves as the answer to highly scalable, accurate retrieval processes. It demonstrates its performance against both FAISS and TF-IDF baselines in this paper. The evaluation process uses a 50,000-document cleaned subset of CORD-19 to analyse four retrieval frameworks: lexical vectors, TF-IDF with dimensional reduction, FAISS Flat indexing and our developed ACG system. ACG creates a hierarchical centroid-based graph through K-Means clustering, which it extends with k-NN centroid graphs for navigation within clusters that employ HNSW indexing. The system automatically increases cluster sizes during search operations through distance threshold management. Under SBERT embedding conditions, ACG demonstrates a perfect performance in queries with Recall@1 and Recall@5 while maintaining retrieval times below 0.43 milliseconds, surpassing FAISS in accuracy while requiring few additional computational resources. Other evaluations included t-SNE cluster visualisation, latency distribution analysis, and precision at different recall values. The combination of performance attributes from ACG makes it an ideal choice for semantic precision and high-throughput operations in RAG pipelines and QA systems.