Autonomous RAG-based Enterprise Search & Analytics: A Hybrid Retrieval Framework for Grounded Question Answering

Abstract

Enterprise organizations increasingly hold their institutional knowledge in scattered, unstructured document repositories that traditional keyword search cannot interpret reliably. This paper presents an Autonomous RAG-based Enterprise Search & Analytics system, a prototype platform that couples hybrid information retrieval with large language model (LLM) generation and a live operational analytics layer. The retrieval stage runs a custom BM25 keyword ranker and a dense vector ranker in parallel, and a configurable fusion layer — Reciprocal Rank Fusion (RRF) or weighted score fusion, switchable at run time — merges the two ranked lists before the top-scoring passages are handed, with citations, to Claude 3.5 Sonnet for grounded answer generation. A companion dashboard exposes real-time key performance indicators so that operators can observe and tune retrieval behaviour without redeploying the system. We report a preliminary pilot evaluation of the prototype over a corpus of ten enterprise policy documents and roughly 30 to 40 test queries: a mean end-to-end latency of 2.0 seconds against a 3-second target, top-1 retrieval scores in the 75 to 95% range, approximately 70% of queries classified as high-confidence, and zero crashes across the test session. The pilot is explicitly small in scale and its numbers are treated as indicative rather than generalizable; the paper is correspondingly transparent about this limitation and about the prototype's current use of placeholder embeddings and an in-memory vector store, both slated for replacement in future iterations.