NovusAI – Advanced Agentic RAG for Drug Repurposing

NovusAI is an advanced Agentic RAG (Retrieval-Augmented Generation) system for drug repurposing and biomedical intelligence.
It orchestrates multiple specialized agents to retrieve evidence from heterogeneous biomedical sources, synthesize insights using a large language model, and generate structured answers with visual analytics.

Unlike traditional RAG, NovusAI uses an agentic orchestration layer where each retriever is an autonomous agent operating on domain-specific sources. Evidence is merged and ranked before synthesis, enabling explainable, multi-hop biomedical reasoning.

The system supports private document ingestion, persistent sessions, and automated report generation.

Live Deployment

• Frontend (Netlify): Live Link
• Backend (Render): https://novusai-backend.onrender.com

Use Case: Accelerating Drug Repurposing

Traditional biomedical research is siloed—patent data, clinical trials, and academic literature rarely talk to each other. NovusAI solves this by acting as an autonomous research assistant that:

• Identifies New Pathways: Cross-references existing FDA-approved drugs with new disease targets found in recent PubMed literature.
• Risk Mitigation: Scans clinical trial failures and patent legalities to assess the feasibility of repurposing.
• Rapid Synthesis: Reduces weeks of manual literature review into seconds of structured, cited intelligence.

Key Features

• Multi-Agent Orchestration: 6+ specialized agents (Patent, Clinical, Literature, Web, Market, Internal) working in parallel.
• Biomedical Entity Intelligence: Integrated with EBI OLS4 API for automated drug/disease entity extraction and synonym expansion.
• Private Knowledge Vault: Secure RAG for internal proprietary documents (.pdf, .txt) using Supabase Vector Storage.
• Visual Analytics Engine: Converts complex evidence data into interactive charts (ReCharts) for market trends and patient outcomes.
• Explainable AI: Every insight includes inline citations and source-linking to the original research or patent filing.
• Contextual Memory: Persistent session states allowing for iterative, multi-turn biomedical discovery.

Screenshot

📸 Product Interface

Main Intelligence Dashboard	JWT Based Login/Signup
Employee Approval Page	Main Research Agent

📂 View Additional Modules (Knowledge Vault & PDF Reports)

Private Knowledge Vault

The secure admin interface for document ingestion and vectorization.

Automated Synthesis Report

Generated high-fidelity biomedical summary.

High-Level Flow (Agentic RAG)

flowchart TB
    U[User Query]
    PS[Pre Synthesis]
    ORCH[Orchestration Layer]

    PAT[Patent Agent - EPO OPS]
    CLIN[Clinical Agent - ClinicalTrials]
    LIT[Literature Agent - PubMed]
    WEB[Web Intelligence - DuckDuckGo]
    MKT[Market Agent]
    INT[Internal Knowledge - Supabase]

    EVID[Evidence Builder]
    SYN[Synthesis Agent - Groq Llama 3.1]
    VIS[Visualization Agent]
    PDF[PDF Agent]
    DB[(Session Database)]

    U --> PS --> ORCH

    ORCH --> PAT
    ORCH --> CLIN
    ORCH --> LIT
    ORCH --> WEB
    ORCH --> MKT
    ORCH --> INT

    PAT --> EVID
    CLIN --> EVID
    LIT --> EVID
    WEB --> EVID
    MKT --> EVID
    INT --> EVID

    EVID --> SYN
    SYN --> VIS
    VIS --> PDF
    SYN --> DB

Loading

Agentic RAG Pipeline

1. Pre-Synthesis Layer

The entry point focuses on linguistic precision and query expansion.

• Entity Extraction: Automatically isolates Disease and Drug entities from natural language queries.
• Synonym Expansion: Connects to the EBI OLS4 API to build comprehensive synonym sets, ensuring the search covers all scientific and trade names.

2. Orchestration & Retrieval

A parallelized agentic layer that queries diverse data silos simultaneously.

Agent	Source	Data Domain
Patent Agent	EPO OPS	Intellectual property, chemical filings, and legal status.
Clinical Agent	ClinicalTrials.gov	Study phases, recruitment status, and primary endpoints.
Literature Agent	PubMed	Academic journals and peer-reviewed clinical research.
Web Intel Agent	DuckDuckGo	Real-time news, press releases, and market alerts.
Market Agent	Mock Data	Commercial trends, pricing, and competitive landscape.
Internal Agent	Supabase	Proprietary documents and historical knowledge.

3. Data Processing & Synthesis

Transforming raw, heterogeneous data into structured intelligence.

Evidence Builder

• Normalization: Standardizes units, dates, and nomenclature across all 6 agents.
• Merging: Deduplicates information and ranks evidence based on source credibility.

LLM Synthesis

• Engine: Groq API
• Model: llama-3.3-70b-versatile
• Output: Generates high-fidelity summaries with inline citations.

4. Analytics & Visualization

The system prepares JSON-ready objects for front-end rendering:

• Market Trends: Historical and projected growth curves.
• Patient Outcomes: Comparative bar charts (Treated vs. Untreated).
• Clinical Roadmap: Pie charts or timelines showing Study Phases (I, II, III, IV).

5. Persistence & Continuity

• Storage: Saves synthesized answers to a permanent database.
• Session State: Enables "rebuild" functionality where the model remembers previous context for iterative discovery.

Private Knowledge Vault (Admin Only)

• File types: .pdf, .txt
• Storage: Supabase (company_docs bucket)
• Used by Internal Knowledge Agent

Authentication

• JWT-based Roles:
• admin → upload documents
• employee → query only

---

Technical Challenges & Solutions

Building a multi-agent system for biomedical data presented several "real-world" hurdles. Below is how I engineered past them:

1. Model Reliability & Constraint Following

• Challenge: Initial testing with local models (Ollama/Phi) resulted in poor instruction following and hallucinated citations.
• Solution: Migrated to Groq (Llama-3.3-70B) and implemented rigorous Prompt Engineering with few-shot examples to ensure the model strictly adheres to biomedical formatting and citation rules.

2. Information Noise vs. Evidence Quality

• Challenge: Agents initially returned a high volume of low-value data, cluttering the synthesis layer.
• Solution: Implemented a Domain-Specific Ranking System.
  • Patents: Ranked by expiry date and legal status.
  • Literature: Ranked by citation count and impact factor.
  • Result: Only the top 5 most authoritative pieces of evidence are passed to the LLM.

3. The "Sparse Search" Problem (Nomenclature)

• Challenge: Queries for a specific drug often failed because different sources used different names (trade names vs. chemical names), leading to near-zero results.
• Solution: Integrated Synonym Expansion via the EBI OLS4 API. The system now automatically expands a single user query into a comprehensive set of scientific and commercial synonyms before the agents begin retrieval.

4. Resolving Contradictory Agent Intelligence

• Challenge: Agents occasionally provided conflicting information (e.g., a web alert claiming a drug failure vs. a patent filing showing active development).
• Solution: Developed a Source-Credibility Scoring Matrix. I assigned weighted priority to data sources:
  • Patent Data (High) > Clinical Trials > Literature > Web Intel (Low).
  • The LLM is prompted to prioritize "Source Truth" based on these weights when synthesizing the final report.

Tech Stack

Backend

• FastAPI
• SQLAlchemy
• JWT
• Supabase Storage
• Groq API

Frontend

• React + Vite
• Mantine UI
• Axios
• ReCharts

Deployment

• Backend → Render
• Frontend → Netlify

Local Setup

Backend

cd backend
pip install -r requirements.txt
uvicorn app.main:app --reload

Frontend

cd frontend
npm install
npm run dev

Environment Variables (Backend)

SUPABASE_URL=xxxxxxxxxx
SUPABASE_SERVICE_KEY=xxxxxxxxxx
SUPABASE_ANON_KEY=xxxxxxxxxx
JWT_SECRET_KEY=xxxxxxxxxx
JWT_ALGORITHM=xxxxxxxxxx
ENV=xxxxxxxxxx
PUBLIC_API_URL=xxxxxxxxxx
GROQ_API_KEY=xxxxxxxxxx
GROQ_BASE_URL=xxxxxxxxxx
MODEL_NAME=xxxxxxxxxx
CONSUMER_KEY=xxxxxxxxxx
CONSUMER_SECRET=xxxxxxxxxx

Environment Variables (Frontend)

VITE_API_BASE_URL=https://<backend-url>

Repository Structure

NovusAI/
  backend/
  frontend/

Author

Devashish Mishra B.Tech | AI/ML | Full-Stack | Cloud