Task-Adaptive LLM Routing Architecture for Multi-Modal Educational Knowledge Retrieval Agents
Year: 2026
Type: conference-paper
Abstract: Undergraduate engineering students frequently encounter challenges in accessing real-time, domain-specific academic support. While Large Language Models (LLMs) offer potential solutions, generic implementations often lack multimodal capabilities (processing text, audio, and visual data) and fail to provide reliable, hallucination-free guidance in specialized fields such as Information Security. This paper proposes “Karn,” a multimodal conversational agent integrated into Telegram. The system utilizes an intelligent model-routing architecture to optimize inference latency and accuracy by selecting appropriate specialized models for distinct query types. To address the reliability issues common in generative models, Karn employs a dual-mode architecture: a dynamic Retrieval-Augmented Generation (RAG) layer powered by MongoDB Atlas Vector Search for user-uploaded documents and a static, zero-latency command-triggered layer (e.g., /wiki) for authoritative, hardcoded Information Security resources. This bifurcated approach guarantees hallucination-free retrieval for core study materials while maintaining adaptive reasoning for dynamic inputs. We conducted a performance analysis focusing on three key indicators: system latency, response precision, and the architecture's ability to scale under load. Our findings confirm that by embedding this tool within a familiar interface, we can provide immediate, context-sensitive academic support without the friction of traditional web platforms.
Institution: Mohan Babu University
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
SDG: Quality education
Open Access status: hybrid
OpenAlex ID: W7167635237