KumbhDoot: LLM-Bounded Architecture for Mass-Gathering Assistants

A recent study presents KumbhDoot, an intelligent assistant for pilgrims at the Nashik Simhastha Kumbh Mela, one of the largest religious events globally. This innovative system tackles the complexities of catering to millions of participants who require urgent, repetitive, multilingual, and critical safety information. Unlike typical conversational agents that depend entirely on large language models (LLMs) for every inquiry, KumbhDoot focuses first on semantic similarity. Generative models are only utilized when this retrieval method fails to yield accurate responses. Central to the system is a 'semantic cache,' which acts as a unified retrieval mechanism. This architecture is designed to be scalable, economical, and resilient, avoiding issues associated with LLMs, such as high costs, slow emergency responses, inaccuracies, and connectivity failures. The research paper, titled 'KumbhDoot: A Scale-Ready, LLM-Bounded Architecture for Mass-Gathering Public-Service Assistants,' can be found on arXiv with the identifier 2608.07520, emphasizing a trend towards hybrid AI systems that merge retrieval techniques with generative models for critical real-world scenarios.