We present a retrieval-augmented chatbot that combines dense retrieval with large language models (LLMs) to answer user queries grounded in external documents [1][2]. Our system ingests domain-specific documents into a vector database and uses an encoder to embed text chunks. For each query, the chatbot performs semantic search to retrieve the top-K relevant passages [3]. These passages are optionally re-ranked by a neural cross-encoder to boost precision [4][5]. The top-ranked chunks are prepended to the user’s query as context, and an LLM generates the final answer. To improve efficiency, we implement caching of retrieval results and answers, and include a fallback snippet-generation mode if the LLM fails. We implement the server in Python using FastAPI and open-source libraries (embedding and vector DB) along with custom code for caching and thread-safe operations. In evaluation, reranking yields significant gains in answer correctness (over 15 percentage points in one study) [5], and the entire pipeline achieves high-quality, faithful answers while remaining efficient. Our contributions include a detailed system design with component and sequence diagrams, implementation techniques (vector store, retriever, re- ranker, LLM), and an evaluation demonstrating the benefits of re-ranking and caching.

 Keywords—Retrieval-Augmented Generation, Chatbot, Information Retrieval, Vector Embeddings, Re-ranking, Large Language Models.