This project implements a personalized, empathetic mental health coach using a fine-tuned version of DeepSeek-7B. It combines Retrieval-Augmented Generation (RAG), LoRA-based fine-tuning, and response reranking with Gemini 2.0 Flash.

• 🔍 Semantic Retrieval: FAISS index + bi-encoder (MiniLM) to retrieve contextually similar mental health Q&A.
• 🧠 LLM Fine-Tuning: DeepSeek-7B fine-tuned with LoRA (on Q_proj, V_proj) using real-world mental health conversations.
• 💡 RAG Pipeline:
  • Generate 3 diverse responses using DeepSeek-LoRA
  • Rerank using Gemini 2.0 Flash (or DeepSeek R1 via Ollama)
• ⚡ 4-Bit Inference: Efficient inference using bitsandbytes with NF4 + double quantization on a 16GB GPU.
• 🔒 Session Encryption: AES-based Fernet encryption used for protecting user data in Streamlit sessions.

• transformers + peft + bitsandbytes
• faiss, sentence-transformers
• streamlit for UI
• cryptography for secure session handling
• google.generativeai for Gemini reranking

Ask a question like:

I'm feeling overwhelmed with work and can't focus.

And get a grounded, personalized response pulled from real-world cases, fine-tuned with empathy.

1. Input → Embedding → FAISS Search
2. Retrieve Top-K Relevant Questions
3. Construct Prompt + Few-Shot Examples
4. Generate 3 Candidates via DeepSeek-7B-LoRA
5. Rerank with Gemini 2.0 Flash
6. Return Best Response to User

Session-level user profile and chat history are encrypted using Fernet (AES 128) for safe interaction and privacy preservation.

/app
├── chatbot_logic.py
├── main.py
├── faiss_index.bin
├── counselchat_with_embeddings.pkl
├── lora-deepseek/
└── final/

• Summarization-based memory compression
• Long-form context (DeepSeek-Prover-V2)
• Real-time user mood tracking and topic memory

Jwalith, MS in Data Science
Stony Brook University