⚕️ReassureAI

A Hybrid AI Healthcare Support System

Bridging Modern Biomedicine with Ancient Ayurvedic Wisdom

ReassureAI provides mental wellness support, medical report simplification, and Ayurvedic health guidance through an intelligent tri-model AI pipeline. Built for the Indian context. Not a medical diagnosis tool. A hybrid AI healthcare assistant bridging modern biomedicine with Ayurvedic wisdom.

---

📋 Table of Contents

• What is ReassureAI
• Core Features
• System Architecture
• Tech Stack
• Models on HuggingFace
• Getting Started
• Demo Login
• Project Structure
• Team
• Disclaimer
• License

---

🧠 What is ReassureAI?

India has 150 million+ people who need mental healthcare — over 80% go untreated. Patients receive medical reports in clinical English they cannot understand. Ayurvedic knowledge, practiced by millions daily, has no intelligent digital interface.

ReassureAI addresses all three gaps in one unified system:

Problem	ReassureAI Solution
Mental health support inaccessible	24/7 empathetic AI conversation with crisis detection
Medical reports unreadable by patients	Plain-language simplification via biomedical LLM
Ayurveda disconnected from digital health	AyurParam model provides Dosha-based traditional guidance

---

✨ Core Features

Core Features

🧠 Mental Wellness Support

• Empathetic conversational support available 24/7
• Semantic crisis detection — understands implied distress, not just keywords
• Automatic guardian alert via email when crisis detected
• CBT-aligned response patterns

📋 Medical Report Simplification

• Upload blood reports, radiology summaries, discharge notes
• Converted to plain language anyone can understand
• Supports PDF, PNG, JPG, JPEG including handwritten documents
• SHA-256 deduplication — same report never processed twice

🌿 Ayurvedic Guidance

• Traditional Dosha-based wellness recommendations
• Integrates with modern medical perspective
• Powered by AyurParam — purpose-built Ayurvedic LLM
• RAG-enhanced to reduce hallucination

🤖 Modern Chatbot UX

• Markdown rendered responses
• Text-to-speech on every response
• Copy, regenerate, like/dislike feedback per message
• Auto-scroll with timestamp on every message
• Multiline input (Shift+Enter for new line)

🛡️ Dual-stage safety screening

• SEC1 — Pre-inference semantic crisis detection via LLM (not keywords)
• SEC2 — Post-inference contradiction check, herb-drug conflict detection
• Medical disclaimer injected in every response
• Guardian alert system for high-risk situations

🔍 RAG-enhanced accuracy

• Hybrid retrieval — FAISS (semantic) + BM25 (keyword) combined
• Reciprocal Rank Fusion for better result ranking
• Reduces LLM hallucination by grounding responses in trusted sources
• Separate knowledge bases for biomedical and Ayurvedic domains

---

🏗️ System Architecture

%%{init: {
  'theme': 'dark',
  'themeVariables': {
    'lineColor': '#BBBBBB',
    'textColor': '#FFFFFF',
    'edgeLabelBackground': '#222222'
  }
}}%%
graph TD
    UserInput["User Input"]
    ModelSelect["Model Select"]
    DisigenNode["Disigen Node"]

    MentalHealth["Mental Health Branch"]
    SemanticGate["Semantic Understanding Layer\n(LLM-based — Mistral Chain 3)"]
    DNode["D-Node\n(SEC1 — Crisis Decision)"]
    MistralChat["Mistral 7B — Chain 3\n(Mental Health Conversation)"]
    RuleBased["Rule Based\n(CALL TOOLS N8N)"]
    EmailSec["Email + Guardian Alert\n+ Security Procedures"]

    PhysicalHealth["Physical Health Branch"]
    QueryIntelligent["Query Intelligent Layer\n(QIL — Mistral Chain 3)"]

    Chain1["OpenBioLLM — Chain 1\n(Biomedical Reasoning)"]
    OtherMedia["Other Media\n(Image, PDF, Docs — OCR)"]
    Chain2["AyurParam — Chain 2\n(Ayurvedic Reasoning)"]
    Chain3QIL["Mistral — Chain 3\n(QIL + Fusion Support)"]

    CallTools["Call Tools + RAG System\n(FAISS + BM25 Hybrid)"]
    QueryFusion["Query Fusion + Response Check\n(SEC2 — All 3 Chains Merge)"]
    Output["Structured Output\n(TTS + Copy + Feedback)"]

    UserInput --> ModelSelect
    ModelSelect --> DisigenNode

    DisigenNode --> MentalHealth
    DisigenNode --> PhysicalHealth

    MentalHealth --> SemanticGate
    SemanticGate -- "Understands implied\nmeaning + emotion" --> DNode
    DNode -- "Safe — proceed" --> MistralChat
    DNode -- "Crisis detected" --> RuleBased
    RuleBased --> EmailSec

    PhysicalHealth --> QueryIntelligent

    QueryIntelligent --> Chain1
    QueryIntelligent --> OtherMedia
    QueryIntelligent --> Chain2
    QueryIntelligent --> Chain3QIL

    Chain1 --> CallTools
    OtherMedia --> CallTools
    Chain2 --> CallTools
    Chain3QIL --> CallTools

    CallTools --> QueryFusion
    QueryFusion --> Output

    style DNode fill:#ff66cc,stroke:#fff,stroke-width:2px,color:#000
    style DisigenNode fill:#9999ff,stroke:#fff,stroke-width:2px,color:#000
    style SemanticGate fill:#ff9944,stroke:#fff,stroke-width:2px,color:#000
    style Chain1 fill:#44aaff,stroke:#fff,stroke-width:2px,color:#000
    style Chain2 fill:#44ffaa,stroke:#fff,stroke-width:2px,color:#000
    style Chain3QIL fill:#ffdd44,stroke:#fff,stroke-width:2px,color:#000
    style MistralChat fill:#ffdd44,stroke:#fff,stroke-width:2px,color:#000
    classDef default fill:#ccff99,stroke:#333,stroke-width:1px,color:#000
    linkStyle default stroke:#FFFFFF,stroke-width:2px

Loading

User Input
    ↓
Disigen Node (dispatcher)
    ├── Mental Health Branch
    │     ↓
    │   Semantic Understanding Layer (Mistral-7B — LLM, not keywords)
    │     ↓
    │   D-Node SEC1 — Crisis Decision
    │     ├── Crisis → n8n → Guardian Alert Email
    │     └── Safe  → Mistral-7B Conversation
    │
    └── Physical Health Branch
          ↓
        Query Intelligence Layer (QIL)
          ↓
        ┌─────────────────────────────────┐
        │ Chain 1: OpenBioLLM-8B          │
        │ Chain 2: AyurParam (Ayurvedic)  │  ← All 3 run concurrently
        │ Chain 3: Mistral-7B (General)   │
        └─────────────────────────────────┘
          ↓
        RAG — FAISS + BM25 Hybrid Retrieval
          ↓
        SEC2 Query Fusion + Post-Safety Review
          ↓
        Structured Dual-Perspective Response

Two Security Checkpoints:

• SEC1 (Pre-inference): LLM semantic analysis before any response model runs
• SEC2 (Post-inference): Contradiction check, herb-drug conflicts, disclaimer injection

System Architecture Node Explanations

Node	Model / Tech	Role
User Input	React frontend	Raw query, file upload
Model Select	UI dropdown	Mental Health / Physical Health / Report
Disigen Node	FastAPI dispatcher	Routes to correct branch
Semantic Understanding Layer	Mistral-7B (Chain 3)	LLM understands true intent, emotion, implied meaning — NOT keywords
D-Node (SEC1)	Decision logic	Receives semantic analysis → crisis/safe binary decision
Mistral (Chain 3) — Mental Health	Mistral-7B Ollama	Empathetic mental health conversation
Rule Based + n8n	n8n workflow	Guardian alert email on crisis
QIL	Mistral-7B (Chain 3)	Intent, urgency, domain scores, query reformulation
Chain 1 — OpenBioLLM	OpenBioLLM-8B HF	Biomedical clinical reasoning
Chain 2 — AyurParam	Aayupahar 3B Ollama	Ayurvedic reasoning (requires RAG context)
Chain 3 — Mistral	Mistral-7B Ollama	QIL, fusion synthesis, general support
Other Media	OCR (pytesseract)	PDF, image, handwritten docs → text
Call Tools + RAG	FAISS + BM25 + LangChain	Hybrid retrieval, tool calls
SEC2 — Query Fusion	Mistral synthesis	Merges all chains, post-safety, disclaimer
Output	React frontend	TTS, copy, feedback (like/dislike), regenerate

---

🛠️ Tech Stack

Frontend

Technology	Purpose
React 18 + Vite	UI framework
Tailwind CSS	Styling
Framer Motion	Animations
react-markdown + remark-gfm	Markdown rendering
Web Speech API	Text-to-speech (browser native)

Backend

Technology	Purpose
FastAPI	REST API (async)
Python 3.11+	Runtime
Motor	Async MongoDB driver
python-jose	JWT authentication
httpx	Async HTTP client

AI Models

Chain	Model	Role
Chain 1	OpenBioLLM-8B	Biomedical clinical reasoning
Chain 2	AyurParam GGUF 3B	Ayurvedic reasoning
Chain 3	Mistral-7B	QIL, semantic gate, fusion, mental health
Embedder	BAAI/bge-m3	RAG document embedding (1024-dim)

Infrastructure

Technology	Purpose
MongoDB	Primary database
Qdrant Cloud	Vector database (RAG)
FAISS	Local vector search fallback
Local disk	File upload storage
n8n	Guardian alert workflow automation
Ollama	Local LLM inference
Docker	As Running Env

---

🤗 Models on HuggingFace

Model	Description	Downloads
AyurParam-GGUF	GGUF conversions (F16 + Q4_K_M) of bharatgenai/AyurParam for local inference via Ollama

---

🚀 Getting Started

Prerequisites

• Python 3.11+
• Node.js 18+
• MongoDB (running locally)
• Ollama (with mistral and aayupahar models pulled)
• WSL2 Ubuntu (if on Windows) — see .agent/codebase/wsl_setup.md

Installation

# Clone the repository
git clone https://github.com/Aaryam-7d6/ReassureAI.git
cd ReassureAI

# Install all dependencies
make install

# Set up environment variables
cp backend/.env.example backend/.env
# Edit backend/.env with your values (see below)

# Seed the test user
make seed

# Start frontend and backend together
make dev

Environment Variables

Visit .env copy once.

Create backend/.env with these values:

# Required — fill these now
MONGODB_URI=mongodb://127.0.0.1:27017/reassureai
JWT_SECRET=your_32_char_secret_here
OLLAMA_BASE_URL=http://127.0.0.1:11434
FRONTEND_URL=http://localhost:5173

# Fill when needed
HUGGINGFACE_API_KEY=
GROQ_API_KEY=
QDRANT_URL=
QDRANT_API_KEY=
N8N_WEBHOOK_URL=
UPLOAD_DIR=data/uploads
MAX_UPLOAD_SIZE_BYTES=10485760

Generate JWT secret:

python3 -c "import secrets; print(secrets.token_hex(64))"

or

openssl rand -hex 64

Pull required Ollama models:

ollama pull mistral
ollama run hf.co/A-Aryam/AyurParam-GGUF:Q4_K_M

Updating soon...

---

🧪 Demo Login

A test user is pre-seeded for demonstration and development:

Email:    test@reassureai.dev
Password: Test@1234!
Full Name: Test User
Guardian Email: guardian@reassureai.dev

This account is for testing only. Do not use for real health queries.

---

📁 Project Structure

Updating soon...

---

👥 Team

Name	Role
Aarya R. Thakar	Originator & Project Lead · System Architecture · AI Pipeline · Research
Ansh B. Patel	Frontend & Backend Development
Darshan B. Kyada	Frontend & Backend Development
Elvis T. Fernandes	Database

Institution: Department of Computer Science & Engineering, PIET, Parul University, Vadodara

Final Year B.Tech Project — Academic Year 2025–27

---

🔬 Research

Research Paper is progress...

---

⚠️ Disclaimer

Warning

ReassureAI is a student research project developed as a Final Year B.Tech project at Parul University. It is experimental and not medically approved, certified, or compliant with HIPAA, DPDP, or any healthcare regulatory framework.

This tool is for educational and informational purposes only.

• Does not provide medical diagnosis
• Does not replace licensed healthcare professionals
• Does not prescribe treatments or medications
• AI responses may be incorrect, incomplete, or outdated
• Not validated by any government body or medical institution
• All responses should be verified with a qualified doctor

Use at your own risk and with full awareness of its limitations. This project explores how AI can be applied to healthcare — it is a proof of concept, not a clinical tool.

---

If you are experiencing a mental health crisis, please contact:

Helpline	Number	Available
iCall	9152987821	Mon–Sat, 8am–10pm
Vandrevala Foundation	1860-2662-345	24/7
MANAS	14416 or 1800-891-4416	24/7

---

📄 License

Licensed under the Apache License 2.0 — see LICENSE for full terms.

© 2025 Aarya R. Thakar, Ansh B. Patel, Darshan B. Kyada, Elvis T. Fernandes

This license is consistent with AyurParam (bharatgenai/AyurParam) which this project integrates, also licensed under Apache 2.0.

---

Built with ❤️ for 150 million Indians who deserve better healthcare access