AI Research Radar

AI Research Radar is a daily AI academic and technology intelligence pipeline. It collects research signals, ranks them for a research-focused reader, generates a Markdown daily report, and renders the latest report as a GitHub Pages page.

Current version: v0.1.0

GitHub Pages: https://georgeorange-crypto.github.io/ai-research-radar/

Default Mode

Default mode: single

Do not use role_pipeline for daily automation unless you understand the API cost. role_pipeline may call multiple models per item and can be expensive. It is not the default mode.

Core Capabilities

• Multi-source AI research collection
• Daily Markdown report generation
• GitHub Pages HTML report generation
• Research-priority ranking for LLM agents, context compression, open-world learning, and model distillation
• Single-model LLM summary mode with OpenAI-compatible API support
• Source health summary
• Benchmark, dataset, and GitHub project sections
• Evergreen classic paper recall section

Quick Start

pip install -r requirements.txt
python run.py

Generated artifacts include:

• report.md
• index.html
• reports/daily/latest.md
• reports/daily/YYYY/MM/YYYY-MM-DD.md
• data/raw/YYYY-MM-DD.jsonl
• data/processed/YYYY-MM-DD.json

GitHub Actions Setup

The scheduled workflow lives at .github/workflows/daily-report.yml. It runs daily at 06:30 Asia/Shanghai and can also be started manually from:

GitHub -> Actions -> Daily AI Research Radar -> Run workflow

Configure production automation with GitHub Actions Secrets and Variables. Do not commit API keys.

Secrets:

• OPENAI_API_KEY
• or fallback provider keys: KIMI_API_KEY, DEEPSEEK_API_KEY, GLM_API_KEY

Variables:

• OPENAI_BASE_URL
• OPENAI_MODEL
• optional provider variables: KIMI_BASE_URL, KIMI_MODEL, DEEPSEEK_BASE_URL, DEEPSEEK_MODEL, GLM_BASE_URL, GLM_MODEL

Recommended daily settings:

MODEL_MODE=single
OPENAI_SUMMARY_BUDGET=3
MAX_OUTPUT_TOKENS=250
MAX_EVIDENCE_CHARS=1600

Recommended low-cost OpenAI-compatible provider example:

OPENAI_BASE_URL=https://api.moonshot.cn/v1
OPENAI_MODEL=moonshot-v1-8k

Cost Control

The default daily path is intentionally cost-safe:

• MODEL_MODE=single
• single-model summaries are capped by OPENAI_SUMMARY_BUDGET
• output length is capped by MAX_OUTPUT_TOKENS
• evidence length is capped by MAX_EVIDENCE_CHARS
• reports show Summary mode, Provider, Model, LLM summary calls, and Last LLM error

role_pipeline is kept only as an experimental advanced mode. It may call DeepSeek, Kimi, GLM, or other role-specific models for one item and can multiply API usage quickly. Do not enable it for the scheduled daily workflow unless the higher cost is intentional.

Current Limitations

• This release is a daily report generator MVP, not a full web application.
• Ranking still needs a feedback loop and may overweight Agent or video diffusion items.
• Context memory papers such as STALE / Q-RAG may need stronger promotion rules.
• GitHub awesome-list repositories may still be noisy.
• Some source parsers may fail silently or return 0 items.

Release Notes

See RELEASE_NOTES.md for the v0.1.0 GitHub Release notes.