MOBStudio: A Local-First Private Framework for Customized Healthcare Conversational Assistants

Large language models (LLMs) and retrieval-augmented generation (RAG) are widely regarded as effective approaches for grounding AI-generated responses in factual medical knowledge. However, their practical adoption in clinical settings is severely hindered by three primary barriers: the inherent safety risks of model hallucinations, data privacy concerns associated with cloud-based AI services, and the advanced technical skills required to configure reliable pipelines. This technical complexity excludes healthcare professionals from the development process, leaving them as passive end-users rather than active creators of tools adapted to their workflows. The objective of this paper is to present MOBStudio, a framework designed to enable healthcare professionals to independently build and validate customized, private RAG systems. MOBStudio combines a modular, local-first architecture that ensures data privacy with a structured workflow for assistant configuration, knowledge ingestion, and interactive evaluation. A core technical originality of this architecture is the decoupling of the assistant's behavioral state from its underlying data infrastructure. Leveraging this decoupled approach, an integrated Configuration Agent guides clinicians during the design process while preserving their control over the assistant's behavior and knowledge sources. Validation was conducted in three distinct clinical domains: traumatology, HIV/AIDS, and ophthalmology, where healthcare professionals independently created conversational assistants based on local institutional knowledge bases. The framework successfully supported the development of domain-specific conversational assistants tailored to individual clinical practices. The technical evaluation demonstrated effective retrieval grounding and contextually relevant responses, and the usability assessment indicated that clinicians perceived the platform as intuitive, practical and suitable for the independent development of clinical conversational assistants. MOBStudio provides a clinician-centered, local-first privacy framework to construct trustworthy RAG-based conversational assistants. By combining AI-assisted configuration with clinician control over knowledge sources and system behavior, the framework lowers the technical barrier to conversational AI development while preserving data privacy and clinical oversight. Its successful validation in traumatology, HIV/AIDS, and ophthalmology demonstrates its applicability to diverse healthcare domains and its potential to support the practical adoption of trustworthy conversational AI in clinical practice.