ARCHITECTING RETAIL-SCALE VECTOR STORE SYSTEMS FOR AGENTIC GENERATIVE AI

Abstract

Abstract

Modern retail-facing generative AI systems increasingly rely on retrievalaugmented generation (RAG) and agentic execution models to deliver accurate, context-aware responses across product catalogs, technical documentation, and operational metadata.At retail scale, these systems must support semantic retrieval over heterogeneous and continuously evolving knowledge while maintaining predictable latency and operational reliability.This paper presents a big data driven architecture for vector store systems designed to support retail-scale agentic generative AI.The architecture integrates batchoriented ingestion pipelines with incremental updates, document-aware chunking, multimodal embedding generation, and scalable approximate nearest-neighbor indexing.Vector stores are treated as core, governed infrastructure components and are accessed by agents through explicit retrieval tools rather than implicit prompt augmentation.Using representative, synthetic benchmarks derived from retail workloads, we analyze ingestion throughput, retrieval latency, and agent-level efficiency across large embedding corpora constructed from product attributes and visually rich documents.