📡 arxiv-tracker

Personal LLM Paper Radar — Automatically fetches the latest arXiv papers daily, summarizes them in Korean via Gemini, and serves them through a clean web UI

Features · Quick Start · Usage · Configuration · Architecture · Claude Code

---

✨ Features

• Auto Fetch — Pulls LLM-related papers from cs.CL / cs.LG / cs.AI categories with keyword filtering
• Korean Summaries — Gemini API summarizes each abstract into 5 structured sections (contribution / method / results / limitations / keywords)
• Web UI — Tailwind-styled page with search, favorites, and paper detail view
• SQLite Storage — Deduplicates papers and tracks summarization status persistently
• Claude Code Integration — One command /daily_papers runs the full fetch → summarize → display pipeline

---

🚀 Quick Start

1. Set up the environment

conda create -n arxiv python=3.11 -y
conda activate arxiv
pip install -r requirements.txt

2. Get a Gemini API key

Google AI Studio — takes 1 minute, free tier is more than enough

cp .env.example .env
# Open .env and fill in GEMINI_API_KEY=

3. Run

python -m src.fetcher      # fetch papers
python -m src.summarizer   # summarize unsummarized papers
uvicorn src.app:app --reload --port 8000

Open http://localhost:8000 in your browser.

---

📖 Usage

CLI

# Fetch papers based on keywords/categories in config.yaml
python -m src.fetcher

# Summarize all pending papers
python -m src.summarizer

On Mondays, set days_back: 3 in config.yaml to include weekend papers.

Web UI

uvicorn src.app:app --reload --port 8000

Route	Description
/	Paper list with search and favorites filter
/paper/{arxiv_id}	Paper detail with Korean summary

Claude Code

Run claude inside the project folder:

/daily_papers          # fetch + summarize + display top 5 papers in one shot
/search_topic RAG      # one-off topic search (not saved to DB)
/my_save               # persist session decisions to CLAUDE.md

---

⚙️ Configuration

Everything is controlled through config.yaml — no code changes needed to add keywords or swap models.

Key	Default	Description
summary_model	gemini-3.1-flash-lite	Gemini model for summarization
max_output_tokens	1024	Max tokens per summary
categories	cs.CL, cs.LG, cs.AI	arXiv categories to search
keywords	LLM, RAG, RLHF …	Title/abstract filter keywords
max_results	50	Max papers fetched per run
days_back	3	How many days back to look

Full keyword list

keywords:
  - "large language model"
  - "LLM"
  - "language model"
  - "in-context learning"
  - "chain-of-thought"
  - "RAG"
  - "retrieval-augmented"
  - "LLM agent"
  - "tool use"
  - "fine-tuning"
  - "instruction tuning"
  - "RLHF"
  - "alignment"
  - "prompt"

---

🏗️ Architecture

arxiv-tracker/
├── src/
│   ├── fetcher.py      # arXiv API → SQLite
│   ├── summarizer.py   # SQLite → Gemini API → Korean summary
│   ├── app.py          # FastAPI web server
│   ├── db.py           # Single SQLite entry point
│   └── templates/      # Jinja2 + Tailwind HTML
├── config.yaml         # Model, keywords, categories
├── .env                # API keys (git-ignored)
└── data/papers.db      # SQLite DB (git-ignored)

Data flow

arXiv API
   │  arxiv client (delay_seconds=3)
   ▼
fetcher.py ──▶ data/papers.db
                    │  unsummarized rows
                    ▼
             summarizer.py ──▶ Gemini API
                    │  Korean summary
                    ▼
             data/papers.db (summary_ko column)
                    │
                    ▼
              app.py (FastAPI)
                    │
                    ▼
             http://localhost:8000

The runtime LLM (Gemini) and the development tool (Claude Code) are intentionally separate — Gemini summarizes papers, Claude Code writes and debugs the code.

---

🤖 Claude Code Integration

This project uses Claude Code as the development tool throughout.

Feature	Location	Role
CLAUDE.md	CLAUDE.md	Injects project rules and conventions into every Claude session
Slash Commands	.claude/commands/	Automates /daily_papers, /search_topic, /my_save
Skills	.claude/skills/paper-summary/	Enforces the 5-section summary format consistently
Subagent	.claude/agents/paper-evaluator.md	Scores each paper 1–10 by relevance to user interests
Hooks	.claude/settings.json	PreCompact auto-backup · Stop completion beep
Auto Memory	~/.claude/projects/…/memory/	Persists user context across sessions

For a detailed walkthrough of each technique, open claude_tech_overview.html in a browser.

---

📦 Dependencies

Package	Purpose
arxiv	arXiv API client
google-genai	Gemini API SDK
fastapi + uvicorn	Web server
jinja2	HTML templating
python-dotenv	.env loading
pyyaml	config.yaml parsing

---

📄 License

MIT © 2025 192cm