mental-health-support-genai: A Multi-Model Integrated AI Support System for Mental Health

This project is an end-to-end mental health AI system integrating three models—Google Gemini, Meta Llama3.3-70B, and Qwen3-32B—enabling zero-shot classification, RAG retrieval, and empathetic response generation. The system achieved an accuracy of 91.75% and an F1 score of 91.56% on a balanced evaluation set of 400 samples. Its technical architecture includes multi-model parallelism, zero-shot prompting, RAG pipeline, and Agentic workflow, providing a reference for AI applications in sensitive domains.

AI applications in the mental health field need to accurately identify crisis signals and respond empathetically. The mental-health-support-genai project builds a complete system integrating three models, enabling zero-shot mental health text classification, retrieval-augmented generation (RAG), and empathetic responses. The system achieved an accuracy of 91.75% and an F1 score of 91.56% on an evaluation set of 400 samples, demonstrating strong classification performance.

1. Multi-model parallel architecture: Gemini as the main classifier, Llama3.3-70B as an alternative, Qwen3-32B to enhance robustness, supporting cross-validation;
2. Zero-shot prompting engineering: Using confidence threshold prompts to guide the model to output results and confidence levels, with Gemini showing the best performance;
3. RAG pipeline: Using ChromaDB vector storage and the gemini-embedding-001 embedding model, indexing 515 mental health knowledge text chunks;
4. Agentic pipeline: Implementing conditional routing via LangGraph—high-severity inputs trigger RAG, while general consultations generate responses directly.

1. Aggressive NLP preprocessing undermines zero-shot LLM classification performance—excessive preprocessing removes emotional marker words and alters semantics, so keeping the original text yields better results;
2. Analysis of over 260,000 real datasets found that crisis posts have 3.3 times more words than non-crisis ones, which can be used as a heuristic rule to assist classification.

1. Sensitive domains require multi-model redundancy design to provide fault tolerance and cross-validation;
2. Well-designed zero-shot prompts can approach supervised learning performance, reducing reliance on annotations;
3. In the LLM era, traditional preprocessing needs to be re-examined—preserving contextual integrity is more important;
4. Combining RAG with Agentic workflow is both accurate and efficient.

Limitations and recommendations:

1. Cannot replace professional services—clear boundaries and referral mechanisms are needed;
2. Models need regular re-evaluation to adapt to language evolution;
3. Currently only supports English—needs to expand to multi-language;
4. Need to design personalized adaptation mechanisms for different user groups.

mental-health-support-genai is a technically solid and clearly structured mental health AI system prototype. It demonstrates the application of cutting-edge technologies such as multi-model integration, RAG, and Agentic workflow, providing valuable experience for LLM applications in sensitive domains. It is a reference project for researchers and developers in AI for Social Good.