📄 Paper Finder

A research paper search and Q&A tool. Search top security and ML conference papers via DBLP, build a personal library, and ask questions over your papers using a local LLM.

---

✨ Features

• 🔍 Search DBLP by keyword
• 🔒 Filter to Top Security 4 — NDSS, S&P, USENIX Security, CCS
• 🧠 Filter to Top ML 4 — NeurIPS, ICML, ICLR, AAAI
• 🌐 Optionally show all other venues
• 📅 Filter by year range
• 📥 Export selected papers as BibTeX (.bib)
• 📚 Personal library with SQLite persistence
• 🔗 Citation graph explorer (references + citations via Semantic Scholar)
• 📄 Fetch open-access PDFs automatically (ArXiv + Semantic Scholar)
• 📎 Upload local PDFs
• 🤖 Ask questions over your papers using Llama 3 (fully local RAG)

---

🚀 Quick Start (Docker)

First run downloads the Llama 3 model (~4.7GB). Subsequent runs are instant.

Requires Docker Desktop.

git clone https://github.com/thusithathilina/PaperFinder.git
cd PaperFinder
docker compose up --build

• App: http://localhost
• API docs: http://localhost:8000/docs

On first run the ollama-init container will pull Llama 3 automatically. You'll see:

Llama 3 model ready

in the logs when it's done. The model is stored in a named Docker volume and persists across rebuilds.

---

🛠️ Local Development

Prerequisites

• Python 3.12+
• Node.js 22.12+
• Ollama installed

Backend

cd backend
python3 -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt
uvicorn main:app --reload

Frontend

cd frontend
npm install
npm run dev

Ollama (separate terminal)

ollama serve
ollama pull llama3   # first time only

Service	URL
App	http://localhost:5173
API	http://localhost:8000
Docs	http://localhost:8000/docs
Ollama	http://localhost:11434

---

🏗️ Architecture

Browser
  └── Nginx :80
        ├── /api/*  →  FastAPI backend :8000
        └── /*      →  React static files

FastAPI backend
  ├── DBLP API          — paper search
  ├── Semantic Scholar  — citation graph + PDF URLs
  ├── SQLite            — library persistence (data/library.db)
  ├── ChromaDB          — vector store (data/chroma/)
  └── Ollama            — local LLM (llama3)

data/
  ├── library.db        — SQLite database
  ├── pdfs/             — downloaded/uploaded PDFs
  └── chroma/           — ChromaDB vector store

Project Structure

PaperFinder/
├── backend/
│   ├── main.py
│   ├── database.py
│   ├── models/
│   │   └── library.py
│   ├── routers/
│   │   ├── papers.py       # DBLP search + BibTeX export
│   │   ├── library.py      # library CRUD
│   │   ├── graph.py        # citation graph
│   │   ├── pdf.py          # PDF fetch/upload/view
│   │   └── rag.py          # RAG ingest + query
│   ├── services/
│   │   ├── dblp.py
│   │   ├── venues.py
│   │   ├── bibtex.py
│   │   ├── semantic_scholar.py
│   │   ├── pdf_fetcher.py
│   │   ├── pdf_parser.py
│   │   ├── vector_store.py
│   │   └── rag.py
│   ├── requirements.txt
│   └── Dockerfile
├── frontend/
│   ├── src/
│   │   ├── App.jsx
│   │   ├── components/
│   │   │   ├── SearchBar.jsx
│   │   │   ├── FilterPanel.jsx
│   │   │   ├── PaperList.jsx
│   │   │   ├── PaperCard.jsx
│   │   │   ├── LibraryPage.jsx
│   │   │   ├── CitationGraph.jsx
│   │   │   ├── GraphTab.jsx
│   │   │   └── RAGTab.jsx
│   │   └── services/
│   │       └── api.js
│   ├── nginx.conf
│   └── Dockerfile
├── data/               # persisted data (gitignored)
│   ├── library.db
│   ├── pdfs/
│   └── chroma/
├── docker-compose.yml
└── README.md

---

🤖 RAG Pipeline

1. Add papers to your library via search
2. Click 🔍 Fetch PDF — auto-fetches open-access PDF (ArXiv, Semantic Scholar)
3. Or 📎 Upload PDF for paywalled papers
4. PDF is automatically parsed, chunked, and embedded into ChromaDB
5. Switch to Q&A tab and ask questions — Llama 3 answers using retrieved context

---

🔌 API Reference

Method	Endpoint	Description
GET	/api/venues	List supported venues
GET	/api/search?q=...	Search papers
POST	/api/export/bibtex	Export BibTeX
GET	/api/library	Get library
POST	/api/library	Add to library
DELETE	/api/library/{key}	Remove from library
POST	/api/pdf/fetch/{key}	Fetch PDF from open access
POST	/api/pdf/upload/{key}	Upload local PDF
GET	/api/pdf/view/{key}	View PDF in browser
POST	/api/graph/expand	Get citation graph connections
GET	/api/rag/status	RAG ingestion status
POST	/api/rag/query	Ask a question

---

🏛️ Supported Venues

Group	Venue	DBLP Key
Security	NDSS	conf/ndss
Security	IEEE S&P	conf/sp
Security	USENIX Security	conf/uss
Security	CCS	conf/ccs
ML	NeurIPS	conf/nips
ML	ICML	conf/icml
ML	ICLR	conf/iclr
ML	AAAI	conf/aaai

---

🧰 Tech Stack

Layer	Technology
Backend	Python, FastAPI, httpx
Database	SQLite via SQLModel
Vector store	ChromaDB
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
LLM	Llama 3 via Ollama (local)
PDF parsing	PyMuPDF
Frontend	React, Vite, Tailwind CSS
Serving	Nginx (reverse proxy + static files)
Containerisation	Docker, Docker Compose
Data sources	DBLP API, Semantic Scholar API

---

⚙️ Configuration

Copy .env.example to .env to customise:

cp .env.example .env

Variable	Default	Description
DB_PATH	data/library.db	Custom SQLite path
DATA_DIR	data	Root data directory
OLLAMA_HOST	http://localhost:11434	Ollama server URL

---

📝 Notes

• DBLP's stream: venue filter doesn't work via their API — results are post-filtered by venue key
• PDF availability depends on open-access status — ArXiv papers are most reliably fetched
• Llama 3 requires ~8GB RAM; use Mistral 7B for lower memory usage (change OLLAMA_MODEL in rag.py)
• Citation graph clicks don't work in Brave browser due to a canvas pointer event bug — use Firefox or Chrome

---

📄 License

MIT