Providers of relief in distress: RAG-based LLMs as situation and intent-aware assistants

Abstract

In high-stress humanitarian and mental health contexts, timely access to accurate, empathetic, and actionable information remains critically limited, especially for at-risk and underserved populations. This work introduces LLooMi, an open-source, retrieval-augmented generation (RAG) conversational agent designed to deliver trustworthy, emotionally attuned, and context-aware support across domains such as mental health crises, housing insecurity, medical emergencies, immigration, and food access. Leveraging large language models (LLMs) with structured prompting, LLooMi reformulates user queries into actionable intents, often implicit, emotionally charged, or vague. It then retrieves and grounds responses in a curated, domain-specific knowledge base, without storing personal user data, aligning with privacy-preserving and ethical AI design principles. LLooMi adopts an intent-aware architecture that adapts its tone, content, and level of detail based on the user's inferred psychological state and informational goals. This step enables delivering fast, directive responses in acute distress scenarios or longer, validation-oriented support when emotional reassurance is needed, emulating key facets of therapeutic communication. By integrating NLP-driven semantic retrieval, structured dialogue memory, and emotionally adaptive generation, LLooMi offers a novel approach to scalable, human-centered digital mental health interventions. Evaluation shows an average answer correctness (AC) of 92.4% and answer relevancy (AR) of 84.9%, with high scores in readability, perceived trust, and ease of use. These results suggest LLooMi's potential as a complementary NLP-based tool for mental health support in digital psychiatry and crisis care.