Reimagine Google Pagerank search for generative information retrieval

Traditional Internet search engines rely heavily on web-link PageRank algorithms and lexical keyword indexing to retrieve and rank web pages. However, the paradigm of modern Information Retrieval (IR) is shifting toward Generative Information Retrieval (GIR), where users expect synthesized, intelligent, direct answers rather than static lists of hyperlinks. While modern dense vector retrieval systems excel at semantic search, they evaluate textual documents as isolated embeddings, lacking structural awareness of logical dependencies, topic hierarchies, and conceptual authority. (Brin, S., & Page, L., 1998) This report presents GraphRAG-Search, an AI-powered next-generation search engine designed for web-scale knowledge bases. The proposed system integrates Large Language Models (LLMs) with graph-based authority algorithms to deliver intelligent retrieval, evaluation, and dynamic answer synthesis. By storing document chunks in a dual vector-graph database architecture (Milvus and Neo4j), the system executes a multi-stage retrieval workflow. For any query, a hybrid retriever extracts top candidates using 384-dimensional dense embeddings (all-MiniLM-L6-v2), builds a local candidate subgraph, and applies a personalized PageRank algorithm to measure structural authority. An iterative, stateful LangGraph agent orchestrates search planning, gap analysis, multi-round refinement, and LLM synthesis.The production-grade full-stack architecture consists of a React UI, a Spring Boot orchestration service, a high-performance Python gRPC query processor, and MinIO object storage. Research has shown that combining vector similarity with graph authority (0.6 S vec + 0.4 S PR) yields statistically significant improvements in retrieval precision and reduces generative hallucination compared to standard dense vector RAG implementations. (Hamzić et al., 2026)