An Intelligent Context-Aware Framework for Academic Question Answering Using Semantic Retrieval and RAG

Abstract

Higher education institutions maintain large collections of academic information distributed across syllabi, regulations, examination instructions, programme documents, notices, handbooks, academic calendars, and departmental records. Although these resources are authoritative, students often need to search multiple documents or contact academic staff to resolve routine questions. Conventional keyword search is sensitive to terminology, while standalone conversational language models may produce fluent answers without reliable access to current institutional information. This paper presents a context-aware academic question-answering architecture that combines semantic retrieval, document metadata, student context, and retrieval-augmented generation (RAG). Unlike a conventional chatbot, the proposed approach treats institutional documents as the primary evidence layer and uses the language model mainly for contextual interpretation and response formulation.
The proposed architecture contains five functional stages: institutional knowledge preparation, query understanding, context-aware retrieval, evidence-constrained answer generation, and response validation. Academic documents are cleaned, segmented, enriched with metadata, embedded into a vector representation, and indexed for retrieval. A student's programme, semester, course, and academic year can be used as optional contextual signals to refine retrieval. Semantic similarity is combined with lexical matching and metadata constraints to identify evidence relevant to the query. Retrieved passages are then supplied to an LLM with explicit instructions to answer from the available evidence and to acknowledge insufficient information when appropriate. The paper defines an experimental methodology using Precision, Recall, F1-score, Recall@K, Mean Reciprocal Rank, answer relevance, faithfulness, latency, and user-oriented usefulness.

Keywords