A Discord bot that generates Traditional Chinese summaries of arXiv papers with PDF reports and provides monthly curated lists of influential research papers.

• 📄 Paper Summarization: Extract arXiv papers and generate structured Traditional Chinese summaries
• 🤖 Multi-Stage LLM Pipeline: Pre-sanitizer → Main Summarizer → Validator for quality assurance
• 📊 PDF Reports: Automatically generate professional PDF reports
• 💾 Smart Caching: Redis-based caching for metadata, summaries, and PDFs (7-30 day TTLs)
• ⏱️ Rate Limiting: Multi-tier rate limiting (default, trusted, admin)
• 📈 Cost Tracking: Token usage and cost estimation per request and daily aggregates
• 🔄 Streaming Updates: Real-time progress updates in Discord

Discord ↔ Bot Service ─┬─ LLM Service (OpenAI/vLLM)
                        ├─ Retriever (arXiv/Semantic Scholar)
                        ├─ PDF Exporter
                        ├─ Redis Cache
                        └─ Scheduler (Monthly Push)

1. Stage A (Pre-Sanitizer): Small model cleans and validates metadata
2. Stage B (Main Summarizer): Primary model generates 繁體中文 structured summary
3. Stage C (Validator): Low-temperature validation of JSON schema and language compliance

• Docker & Docker Compose
• Discord Bot Token (Create one here)
• OpenAI API Key (or self-hosted vLLM)

1. Clone the repository

   git clone <repository-url>
   cd discord-arxiv

2. Configure environment

   cp .env.example .env
   # Edit .env with your credentials

3. Required environment variables

   DISCORD_TOKEN=your_discord_bot_token
   OPENAI_API_KEY=your_openai_api_key
   OPENAI_BASE_URL=https://api.openai.com/v1  # or vLLM endpoint

4. Start services with Docker Compose

   docker-compose up -d

5. Check logs

   docker-compose logs -f bot

1. Install dependencies

   pip install -r requirements.txt

2. Start Redis

   docker run -d -p 6379:6379 redis:7-alpine

3. Run the bot

   python -m src.bot.main

Generate a Traditional Chinese summary of an arXiv paper.

Examples:

/summarize 2401.01234
/summarize https://arxiv.org/abs/2401.01234
/summarize https://arxiv.org/pdf/2401.01234.pdf

Output:

• Structured embed with 4 sections (簡介/背景/方法/結論)
• Bullet points highlighting key contributions
• PDF report attachment

Processing time: Typically 10-30 seconds

Get top cited papers from arXiv and conferences (coming soon).

Examples:

/top-papers
/top-papers 2024-11
/top-papers 2024-11 LLM

See .env.example for all available options. Key settings:

Configured per user tier:

• Default: 3 requests/min, 20/day
• Trusted: 6 requests/min, 100/day
• Admin: Unlimited (tracked)

Adjust in .env:

RATE_LIMIT_DEFAULT_PER_MIN=3
RATE_LIMIT_DEFAULT_PER_DAY=20

Uncomment the vllm service in docker-compose.yml and set:

OPENAI_BASE_URL=http://vllm:8000/v1
OPENAI_API_KEY=dummy  # vLLM doesn't require real key

src/
├── bot/              # Discord bot & commands
│   ├── commands/     # Slash commands (/summarize, /top-papers)
│   └── main.py       # Bot entry point
├── llm/              # LLM pipeline
│   ├── client.py     # OpenAI-compatible client
│   ├── pipeline.py   # Multi-stage pipeline (A/B/C)
│   ├── prompts/      # Prompt templates
│   └── validators/   # Output validators
├── retriever/        # Data retrievers
│   ├── arxiv.py      # arXiv API client
│   └── semantic_scholar.py  # Citation data
├── exporter/         # PDF generation
│   └── pdf.py        # ReportLab PDF exporter
├── config/           # Configuration
│   ├── settings.py   # Pydantic settings
│   ├── logging.py    # Structured logging
│   └── cache.py      # Redis cache layer
└── telemetry/        # Observability (future)

Redis key namespace: dra:*

The bot tracks token usage and estimated costs:

• Per-request tracking: tokens_in, tokens_out, cost_estimated, duration_ms
• Daily aggregates stored in Redis
• Simplified pricing estimates (configurable in llm/client.py)

View daily costs (requires admin access):

from src.config.cache import RedisCache
cache = RedisCache("redis://localhost:6379/0")
await cache.connect()
costs = await cache.get_daily_cost("2024-11-22")
print(costs)  # {'tokens_in': 12500, 'tokens_out': 3200, 'cost_estimated': 0.025}

pytest tests/

• See CLAUDE.md for architecture guidelines
• See TDS.md for complete technical design specification
• See AGENDA.md for development phases and milestones

1. Check bot is online: docker-compose logs bot
2. Verify SYNC_COMMANDS=true in .env
3. For guild commands, set COMMAND_SCOPE=guild and COMMAND_GUILD_IDS
4. Manually sync: Bot will sync on startup if configured

# Check Redis is running
docker-compose ps redis
# Check logs
docker-compose logs redis

Chinese characters require proper font support. The PDF exporter uses reportlab's default Unicode support. For custom fonts, add TTF files and register in src/exporter/pdf.py.

• Check OPENAI_BASE_URL and OPENAI_API_KEY
• Verify API quota and rate limits
• Check logs: docker-compose logs bot | grep llm_error

See AGENDA.md for detailed phases:

• ✅ Phase 1: Foundations (Bot, Retriever, LLM Pipeline)
• ✅ Phase 2: Feature Completion (/summarize command, PDF export)
• 🚧 Phase 3: Hardening (Tests, error handling, load testing)
• 📋 Phase 4: Launch (Monthly scheduler, /top-papers, observability)

This project follows the Technical Design Spec in TDS.md. Key principles:

• Multi-stage LLM pipeline for quality
• Aggressive caching for cost control
• Structured logging for observability
• Traditional Chinese (繁體中文) output
• Docker-first deployment

[Add your license here]

For issues and questions:

• Check CLAUDE.md for development guidelines
• Review TDS.md for technical specifications
• See structured logs: docker-compose logs bot | jq