First-Aid Buddy: An Intelligent First Aid Assistant System Based on Multi-Agent Architecture

First-Aid Buddy is a multi-agent first aid assistance system integrating large language models, RAG (Retrieval-Augmented Generation), and real-time APIs. Its core innovation lies in the dual-agent architecture of triage assessment and first aid instruction, which can provide users with localized and multilingual emergency medical guidance. It is positioned as an auxiliary tool for professional medical help rather than a substitute.

In sudden medical emergencies, the general public often lacks professional first aid knowledge and may panic in critical moments. First-Aid Buddy was developed to address this pain point, using AI technology to convert professional first aid knowledge into actionable step-by-step guidance. The project is built with Python and Streamlit, with core innovation in the dual-agent architecture (triage assessment + first aid instruction). It integrates geolocation services, video tutorial retrieval, and multilingual translation functions to achieve "anytime, anywhere, personalized" first aid support.

The system is divided into user interaction layer, intelligent agent layer, and external service layer. Core dual-agent collaboration mechanism:

• Triage Agent: Assesses symptom severity on a 1-5 scale based on ESI standards, and determines subsequent processes (self-guided instructions for minor cases / advice to call emergency services + nearby hospital information for severe cases);
• First Aid Agent: Executes RAG retrieval of St. John Ambulance manuals, searches for YouTube-certified first aid videos, and queries nearby medical institution information in parallel. Technology stack selection: Frontend Streamlit; Large model: Groq Cloud's llama-3.3-70b-versatile; Vector retrieval: FAISS + all-MiniLM-L6-v2; Keyword retrieval: BM25; Geolocation services: OpenStreetMap's Overpass API and Nominatim.

• Intelligent Triage Assessment: Classifies based on ESI international standards to ensure recommendations match the situation;
• RAG-based Authoritative Knowledge Retrieval: Uses St. John Ambulance manuals as the knowledge base, retrieves via vector + BM25 indexing to avoid large model hallucinations;
• Real-time Geolocation Services: For cases with ESI score >2, calls location to query nearby hospitals and displays local emergency numbers (supports multiple countries);
• Multilingual Support: Automatically detects and translates user language to lower usage barriers;
• Video Tutorial Retrieval: Integrates YouTube Data API to obtain visual first aid tutorials from certified channels.

The knowledge base includes ESI triage manuals (for training the triage agent) and St. John Ambulance first aid manuals (as RAG retrieval sources). When first launched, it automatically processes PDF documents: extracts text (using OCR if necessary), builds FAISS vector indexes and BM25 keyword indexes, and stores them persistently to improve subsequent startup speed.

Application Scenarios: Home first aid (handling cuts/scalds, etc.), outdoor emergencies (when far from medical institutions), travel assistance (localized first aid information), first aid training (auxiliary learning). Social value: Provides timely guidance before professional medical help arrives. It is clearly positioned as an auxiliary tool; in critical situations, local emergency numbers should be dialed.

Deployment Steps:

1. Clone the code repository and install Python dependencies;
2. Install Tesseract OCR engine;
3. Configure API keys (Groq/YouTube/Serper, etc.);
4. Place PDF knowledge base in the specified directory;
5. Run the Streamlit application. Building indexes takes time during the first launch; subsequent launches use pre-built indexes directly for fast response.

The project demonstrates the potential of multi-agent architecture in vertical domains, achieving precise and reliable services through task decomposition; RAG technology solves the problem of large model hallucinations in professional fields. This model can be extended to legal consultation, technical support, and other fields; in the future, image recognition capabilities can be integrated to obtain more accurate assessment guidance by uploading photos of injuries.