🔬 IRPA : Intelligent Research Paper Assistant

An Agentic RAG system for grounded research paper Q&A, real-time claim verification against modern literature, and multi-session literature review.

🚀 Explore Live Demo • 📖 Overview • ✨ Key Features • 🏛️ Architecture • 🛠️ Tech Stack • ⚡ Quickstart

---

💡 Why IRPA Exists

In AI & Machine Learning, research moves at breakneck speed. A foundational claim or benchmark established in a 2023 paper can be superseded within months by novel architectures and optimization techniques.

Standard RAG chatbots merely tell you what a paper argued they cannot tell you whether that argument still holds up today.

IRPA treats "Is this claim still true?" as a first class citizen. Instead of standard single hop retrieval, IRPA leverages an autonomous LangGraph agent loop to dynamically route queries, challenge assertions across live scientific publications via web search, detect retrieval degradation, and self correct on the fly.

---

✨ Key Features

Capability	Description
◫ Multi-Source Ingestion	Ingest local PDFs, Markdown, TXT, live web URLs, or directly pull papers from arXiv via paper ID / title.
◈ Autonomous Routing Loop	Agentic router categorizes intents into Direct Answer, Vector Store Retrieval, or Claim Verification.
✓ Live Claim Verification	Validates factual claims against recent literature via Tavily web search, highlighting superseded findings with direct citations.
↺ Self Correcting Retrieval	A relevancy grading node detects poor chunks, triggering automated query rewrites and retries to avoid hallucinations.
⊞ Isolated Multi Session Memory	Chat sessions maintain independent Qdrant vector collections and SQLite state checkpointers preventing context bleed across papers.
↳ /btw Side Channel	Ask quick, off-topic side queries without corrupting your active research session's conversational memory.
🗲 Disk Cached Embeddings	Content chunks are SHA-256 hashed and cached locally to prevent redundant, costly embedding API re calls.
✦ Auto Named Threads	Context-aware LLM generates concise, descriptive session titles from your initial query.

---

🏛️ System Architecture

                                ┌───────────────────────────┐
                                │   User / Streamlit UI     │
                                └─────────────┬─────────────┘
                                              │
                                              ▼
                                ┌───────────────────────────┐
                                │   LangGraph Agent Router  │
                                └──────┬──────┬──────┬──────┘
                                       │      │      │
             ┌─────────────────────────┘      │      └────────────────────────┐
             ▼                                ▼                               ▼
   ┌────────────────────┐          ┌────────────────────┐          ┌────────────────────┐
   │ Direct Answer Node │          │  Retrieval Agent   │          │ Claim Verification │
   └────────────────────┘          └──────────┬─────────┘          └──────────┬─────────┘
                                              │                               │
                                    ┌─────────┴─────────┐                     ▼
                                    ▼                   ▼           ┌───────────────────┐
                          ┌──────────────────┐ ┌──────────────────┐ │ Tavily Web Search │
                          │  Qdrant Vector   │ │ Tavily Search    │ └───────────────────┘
                          │   Collection     │ │ (Web Context)    │
                          └─────────┬────────┘ └──────────────────┘
                                    │
                                    ▼
                         [ Relevancy Evaluation ]
                                    │
                       ┌────────────┴────────────┐
                       │ (Pass)                  │ (Fail / Low Confidence)
                       ▼                         ▼
            ┌─────────────────────┐   ┌───────────────────────┐
            │   Generate Final    │   │ Query Rewrite & Retry │
            │ Grounded Answer     │   └──────────┬────────────┘
            └─────────────────────┘              │
                       ▲                         │
                       └─────────────────────────┘

State Persistence: Conversational history, agent state checkpoints, and session metadata are persisted through an embedded SQLite checkpoint database alongside isolated Qdrant vector spaces.

---

🛠️ Tech Stack

Layer	Technology	Rationale & Highlights
Agentic Framework	LangGraph	Explicit cyclical graph state machine with dynamic routing, retries, and fallback handling.
LLM & Embeddings	Google Gemini	High throughput gemini-3.5-flash for reasoning paired with native text embeddings.
Vector Database	Qdrant Cloud	Low latency vector search with per session collection isolation and payload filtering.
Live Web Retrieval	Tavily Search API	Optimized for agent workflows to discover recent counter evidence and fresh research.
Evaluation Suite	DeepEval	LLM assisted evaluation for Faithfulness, Answer Relevancy, and Contextual Precision.
Frontend	Streamlit	Fast, responsive chat UI featuring word by word streaming and instant access to verified sources.
Container & Cloud	Docker + AWS EC2	Fully containerized environment deployed on an AWS EC2 instance.

---

⚡ Local Installation & Setup

Prerequisites

• Python 3.10+
• Git
• Qdrant Cloud Cluster & Google AI Studio API Key

1. Clone the Repository

git clone https://github.com/octorohan/IRPA.git
cd IRPA

2. Environment Setup

python -m venv venv
# On Windows:
.\venv\Scripts\activate
# On Linux/macOS:
source venv/bin/activate

pip install -r requirements.txt

3. Configure Credentials

Create a .env file in the root directory:

GOOGLE_API_KEY="your_google_ai_studio_key"
TAVILY_API_KEY="your_tavily_search_key"
QDRANT_URL="https://your-cluster-id.eu-central-1-0.aws.cloud.qdrant.io"
QDRANT_API_KEY="your_qdrant_api_key"
APP_PASSWORD="your_secure_password"

4. Launch the Application

streamlit run app.py

Open your browser at http://localhost:8501.

---

🐳 Running with Docker

Option A: Pull from Docker Hub (Recommended)

# 1. Create host storage directories
mkdir -p ~/irpa-data/embedding_cache
touch ~/irpa-data/sessions.json
touch ~/irpa-data/checkpoints.db

# 2. Run container
docker run -d \
  -p 8501:8501 \
  --env-file .env \
  -v ~/irpa-data/embedding_cache:/app/embedding_cache \
  -v ~/irpa-data/sessions.json:/app/sessions.json \
  -v ~/irpa-data/checkpoints.db:/app/checkpoints.db \
  --restart unless-stopped \
  --name irpa-app \
  octorohan/irpa-app:latest

Option B: Build Locally

docker build -t irpa-app .
docker run -d -p 8501:8501 --env-file .env --name irpa-app irpa-app

---

📊 Evaluation & Benchmarking

IRPA integrates DeepEval to quantitatively grade RAG pipelines using LLM-as-a-Judge against synthetic golden datasets generated from ingested research papers:

python evaluate.py

• Faithfulness Metric: Ensures answers are strictly grounded in retrieved research context.
• Answer Relevancy: Measures whether responses directly address user questions.
• Contextual Precision & Recall: Validates that retrieved chunks are dense in signal and noise-free.

---

📂 Project Structure

IRPA/
├── app.py                  # Streamlit application entry point & UI
├── backend/
│   ├── vector_store.py     # Qdrant client, isolated collections & disk-caching
│   ├── rag_graph.py        # LangGraph state machine, agent nodes & dynamic router
│   ├── paper_loader.py     # Document loaders (PDF, Web scraping, arXiv API)
│   ├── btw_handler.py      # /btw side-channel conversational memory logic
│   └── models.py           # Pydantic data schemas & state models
├── documents/               # Sample research benchmarks & papers
├── evaluate.py              # DeepEval testing harness
├── Dockerfile               # Production container image specification
├── DOCKER_GUIDE.md          # Comprehensive AWS EC2 & Docker deployment guide
├── requirements.txt         # Project dependencies
└── README.md

---

📌 Known Considerations

• Quota Throttling: Running on shared API free tiers may hit rate limits during heavy concurrent chunking.
• Qdrant Cloud Inactivity: Free-tier clusters may hibernate after extended inactivity.
• Session Teardown: Deleting a session from the UI currently retains historical checkpoint entries in SQLite.

---

👨‍💻 Author & Contact

Built with ❤️ by Rohan Datusalia

---

⭐ If you find this project helpful, please consider starring the repository!