Scholar Inbox: Personalized Paper Recommendations for Scientists

This project is a high-fidelity replication (~96%) of the research paper "Scholar Inbox: Personalized Paper Recommendations for Scientists" (arXiv:2504.08385v2).

It provides a local-first paper recommendation platform that collects data from arXiv, visualizes papers in a 2D embedding space with t-SNE, and learns user preferences with a weighted logistic regression model.

---

Key Features

1. Daily Digest

   • Ranks papers by relevance on a score scale of $[-100, 100]$.
   • Highlights the most important contribution or idea sentences in each abstract for speed reading.
   • View PDF & AI Chat Assistant (Neural RAG): opens the official arXiv PDF inside the app and provides a RAG chat assistant with hierarchical Parent-Child Chunking. Semantic retrieval uses Google text-embedding-004 (768-dimensional) with batching and disk caching, with local TF-IDF fallback.
   • Suggests similar papers using real-time cosine similarity.
   • Supports BibTeX preview and one-click reference copy.

2. Scholar Maps

   • Visualizes the paper collection in 2D with t-SNE.
   • High-performance Canvas map with drag-to-pan and scroll-to-zoom.
   • Shows dynamic hierarchical topic labels based on zoom level (broad field → subfield → paper title).
   • Diversity Active Learning (K-Means): runs K-Means clustering ($k=5$) on 20 uncertainty-boundary papers (probability $\approx 0.5$) to recommend 5 papers from 5 different research groups, helping the model learn broader user preferences faster.

3. Collections & Search

   • Searches papers using keywords or free-form text (semantic and lexical search).
   • Lets users create personal paper folders.
   • Suggests collection expansion based on the average vector of items inside a folder.

4. Conference Planner

   • Ranks poster sessions by average relevance to user preferences.
   • Adds posters to a detailed personal schedule with quick links for PDF reading and AI RAG Q&A.

5. AI Research Workspace

   • LangGraph Multi-Agent Workflows: sets up an automated agent pipeline (ArxivSearchAgent → parallel PaperCriticAgent workers → LiteratureReviewAgent).
   • Phased Technical Roadmap: generates a 3-stage Vietnamese research plan (baseline replication → module integration → evaluation), milestones, and hardware/software risk tracking.
   • Custom Markdown-to-JSX Compiler: renders interactive markdown in the UI and builds dynamic HTML/CSS glassmorphism tables.
   • Hybrid API Router (Gemini & Groq): automatically routes API keys to fallback services for high availability.

---

Technology Stack

• Backend: Python 3.12, Flask, Flask-CORS, LangGraph, NumPy, scikit-learn (TfidfVectorizer, LogisticRegression, t-SNE), Google Gemini API (embed-004), Groq Llama 3.3.
• Frontend: React 19 (Vite), Lucide React, HTML5 Canvas.
• Containerization: Docker, Docker Compose.

---

How to Run

Option 1: Docker (Recommended)

Run the following command from the project root:

docker compose up --build

This starts both services automatically:

• Frontend SPA: http://localhost:5173/
• Backend API: http://localhost:5001/

---

Option 2: Run Locally

1. Start the Backend API (Flask)

Run the backend with Python 3.12:

python backend/app.py

On the first run, the system will fetch the latest 200 papers from the arXiv API and compute t-SNE coordinates.

2. Start the Frontend (React + Vite)

Open a new terminal in the project directory and run:

npm run dev

3. Open the App

Visit the app in your browser at http://localhost:5173/

---

Technical Documentation

For architecture details, code structure, data flow, and the weighted learning math used in the paper, see: TECHNICAL.md