Architectural Design and Internal Working of Health Talk: An Offline-Capable, Multilingual AI Conversational Agent for Clinical Triage and Healthcare Facility Discovery in Regional India

In developing nations such as India, primary healthcare triage faces acute structural bottlenecks, including severe doctor-to-population deficits in semi-urban and rural agglomerations, fragmented municipal health informatics, and linguistic barriers across heterogeneous regional populations. While modern conversational agents and large language models (LLMs) offer promising interactive capabilities, their direct clinical deployment is severely hampered by pharmaceutical hallucination risks, non-deterministic reasoning trajectories, prohibitive cloud API costs, and an absence of localized Indic language models. To resolve this critical dilemma, this paper presents the complete architectural design, algorithmic foundations, and empirical validation of Health Talk AI—a novel, offline-first, multilingual conversational triage agent specifically engineered for Tier-2 regional healthcare ecosystems such as the Greater Nashik Metropolitan Area in Maharashtra, India. Health Talk decouples clinical urgency categorization from generative text synthesis by introducing an inviolable, deterministic three-tier triage rule engine that executes prior to any natural language parsing. The architecture synthesizes: (1) an emergency finite-state pre-filter for instant red-flag detection; (2) a multilingual Natural Language Processing (NLP) subsystem featuring Devanagari script tokenization, Hindi versus Marathi marker-word frequency ratio disambiguation, and RapidFuzz partial-ratio entity extraction for pain intensity (1–10 scale), temporal duration, and anatomical body parts; (3) an offline-first Retrieval-Augmented Generation (RAG) vector engine utilizing dense multilingual sentence embeddings (paraphrase-multilingual-MiniLM-L12-v2) coupled with an inviolable cosine dot-product threshold (tau = 0.45) that mathematically eliminates medical hallucinations; (4) a privacy-preserving geospatial discovery subsystem indexing 347 verified healthcare facilities across Nashik using the spherical Haversine metric with two-decimal-place differential coordinate rounding (~1.1 km); and (5) automated state healthcare entitlement checklists (Ayushman Bharat PM-JAY and Mahatma Jyotirao Phule Jan Arogya Yojana - MJPJAY) paired with multi-lingual emergency first-aid response cards. Evaluated across 91 rigorous automated unit tests, Health Talk achieves a 100% test pass rate, 100% Top-1 retrieval precision on clinical queries in English, Hindi, and Marathi, and an end-to-end execution latency under 28 ms on commodity quad-core CPUs without GPU hardware. This work establishes an authoritative, open-access, zero-cloud-cost framework for safe, equitable, and privacy-compliant digital clinical triage.