Collaborative Multi-Agent Architecture for an Autonomous AI Teaching Assistant with Retrieval-Augmented Pedagogical Workflow

Abstract

Educational technology demands AI systems capable of delivering personalized, accurate, and pedagogically structured learning experiences. Single Large Language Model (LLM) deployments have demonstrated persistent limitations, including restricted pedagogical specialization, inconsistent assessment generation, and susceptibility to factual inaccuracies arising from hallucination. This paper presents the design and implementation of an autonomous AI teaching assistant based on a five-agent collaborative architecture comprising a Query Understanding Agent, Retrieval Agent, Teaching Agent, Quiz Generation Agent, and Feedback and Evaluation Agent, orchestrated through LangGraph—a graph-based stateful workflow framework. A Retrieval-Augmented Generation (RAG) pipeline, grounded in ChromaDB and FAISS vector stores, constrains agent outputs to verified educational content, thereby substantially reducing the incidence of factually unsupported responses. Evaluation of the proposed system against a single-LLM baseline, conducted using SQuAD 2.0 educational subsets and a repository of over 1,000 multiple-choice questions drawn from fifteen Machine Learning topics, demonstrates a task performance improvement of approximately 30 to 40 percent over the baseline, with factual response accuracy exceeding 85 percent. The prototype, developed using Python, LangChain, LangGraph, and locally hosted Llama3 models via Ollama, indicates that specialized multi-agent collaboration can yield measurable gains in output reliability, pedagogical coherence, and contextual adaptability in digital tutoring environments.