VECTRA – Health Assistant
A Safety-Aware Retrieval-Augmented Healthcare AI System
Overview
VECTRA is an AI-powered healthcare decision-support system designed to demonstrate the responsible integration of traditional Machine Learning and Large Language Models (LLMs) in sensitive medical domains.
Unlike conversational symptom checkers or chatbot-style medical assistants, VECTRA follows a guided, guarded, and explainable architecture. The system delivers structured, professional, and non-diagnostic health insights, making it suitable for academic, research, and demonstration purposes.
VECTRA is not a diagnostic tool, not an autonomous agent, and not a replacement for healthcare professionals.
Problem Statement
Many AI-based healthcare assistants rely heavily on LLMs, leading to:
Overconfident or hallucinated medical advice
Poor explainability of predictions
Unsafe automation in high-risk domains
Conversational outputs unsuitable for clinical settings
Pure LLM-driven systems behave like chatbots rather than decision-support tools, which reduces trust and increases risk in healthcare applications.
Proposed Solution
VECTRA introduces a safety-aware hybrid intelligence architecture that:
Separates deterministic prediction from generative explanation
Uses Machine Learning as the primary reasoning engine
Employs Retrieval-Augmented Generation (RAG) in a controlled manner
Enforces strict guardrails at architectural, prompt, and output levels
Degrades gracefully when LLM services are unavailable
The system prioritizes reliability, explainability, and safety over conversational flexibility.
Project Objectives
Demonstrate responsible AI design in healthcare
Maintain professional, clinical output tone
Prevent over-reliance on LLMs
Provide explainable and consistent results
Enforce safety and ethical guardrails by design
Support academic reproducibility and industry demos
High-Level Architecture
End-to-End Flow
User-Provided Symptoms ↓ Symptom Encoding & Validation ↓ Machine Learning Prediction Layer ↓ Semantic Knowledge Retrieval (Vector Database) ↓ Guided RAG Explanation Layer (Optional LLM) ↓ Response Sanitization & Safety Checks ↓ Structured, Professional Output
Core Components
- Machine Learning Prediction Layer
Model: Random Forest Classifier
Input: Structured symptom indicators
Output:
Probabilistic condition predictions
Confidence-weighted results
Purpose
Fast, deterministic inference
Explainable feature-based reasoning
Eliminates hallucinations in core medical reasoning
The ML layer remains independent of LLM availability.
- Knowledge Base & Vector Store
Curated symptom–disease medical documents
Text embeddings stored in Pinecone
Semantic retrieval ensures context relevance
Why Pinecone
Production-ready vector database
Low-latency semantic search
Scalable and reliable
This layer ensures all explanations are grounded in verified knowledge.
- Guided and Guarded RAG Pipeline
VECTRA uses controlled RAG, not free-form generation.
Retrieval
Semantic search retrieves top relevant documents
Context size strictly limited
Prompt Guidance Prompts enforce:
Clinical and educational tone
No conversational language
No emojis or markdown
No diagnosis, treatment, or medication advice
LLM Role
Used only for explanation coherence
Cannot introduce new medical facts
Fully constrained by retrieved context
- Fail-Safe & Fallback Mechanism
If:
OPENAI_API_KEY is missing
LLM service is unavailable
LLM call fails
Then:
System falls back to retrieval-only responses
No crashes or broken outputs
Full functionality retained
This ensures system reliability and reproducibility.
- Response Sanitization & Safety Layer
Before output delivery:
Responses are checked for unsafe claims
Tone and structure are validated
Consistency with ML predictions is enforced
This aligns outputs with clinical communication standards.
Guardrails & Safety Design
VECTRA enforces multi-layer guardrails:
Architectural Guardrails
No autonomous diagnosis
Decision-support only
ML Guardrails
Deterministic, explainable predictions
No stochastic reasoning
Retrieval Guardrails
Curated, closed knowledge base
No open internet access
Prompt Guardrails
Strict clinical language enforcement
No conversational or chatbot tone
Output Guardrails
Structured, professional format
No emojis or markdown
Fail-Safe Guardrails
LLM optionality
Graceful degradation
These guardrails make VECTRA aligned with Responsible AI principles.
Key Features
Hybrid ML + RAG architecture
Safety-aware design
Explainable predictions
Professional, clinical outputs
LLM-optional execution
Fail-safe fallback behavior
Academic and enterprise ready
Dataset & Ethics
Public, non-PHI symptom–disease datasets
No personal or sensitive patient data
Ethical and reproducible usage
Technology Stack Backend
Python 3.9+
FastAPI
Scikit-learn
Pinecone
OpenAI API (optional)
Frontend
Node.js 16+
React / Vite
Form-based UI (non-chat)
Installation & Setup Clone Repository git clone https://github.com/your-username/vectra.git cd vectra
Backend Setup python -m venv venv venv\Scripts\activate pip install -r requirements.txt
Environment Configuration
Create a .env file:
PINECONE_API_KEY=your_pinecone_key PINECONE_INDEX_NAME=vectra-health OPENAI_API_KEY=optional_openai_key
Knowledge Base Indexing python knowledge_base/build_kb.py python vector_store/indexer.py
Frontend Setup cd ui/frontend npm install npm run dev
Running the Application Backend python -m uvicorn api.vectra_api:app --reload
Frontend npm run dev
Assumptions & Limitations Assumptions
Symptoms are self-reported
Used only for educational purposes
Limitations
Not a diagnostic system
Not an emergency tool
No treatment recommendations
These constraints are explicit by design.
Innovation Highlights
Safety-first RAG, not blind LLM usage
Deterministic ML core
Guardrails at every layer
Professional healthcare-grade outputs
Future Scope
Confidence-aware routing (Agentic upgrade)
Severity estimation
Telemedicine triage support
Public health analytics
Regional disease trend analysis
Disclaimer
VECTRA is intended strictly for educational, academic, and research purposes. It does not provide medical diagnosis, treatment, or emergency guidance. Users must consult qualified healthcare professionals for medical decisions.