Research Agent Toolkit is a small Python toolkit for scheduled literature and model-update monitoring with GitHub Actions.
The first preset comes from a PET/MRI research workflow: weekly monitoring for MRI-to-PET, Tau PET, Alzheimer's disease, medical vision-language models, medical CLIP-style models, GitHub releases, and Hugging Face model cards.
The v1.0 scope is deliberately narrow: make one weekly monitoring workflow reproducible before adding broader presets or dashboard features.
简体中文说明见 README.zh-CN.md.
| Capability | v1.0 status |
|---|---|
| Scheduled literature monitoring | Supported |
| NeuroPET / MRI-to-PET / Tau PET / AD preset | Supported |
| Medical VLM / medical CLIP / foundation-model preset | Supported |
| GitHub Actions automation | Supported |
| Chinese weekly report | Supported |
| Model-assisted report writing | Supported |
| Optional report delivery | Supported, off by default |
| Notion workflow | Planned, not in v1.0 |
See the v1.0 completeness audit, architecture note, and release checklist.
This repository started from a practical research need: checking new PET/MRI papers, medical imaging model releases, and related code updates every week, then writing an auditable report from the retrieved metadata.
The toolkit is most useful for:
- biomedical engineering students;
- medical imaging researchers;
- PET (Positron Emission Tomography) / MRI (Magnetic Resonance Imaging) researchers;
- AI-for-science users who want scheduled literature digests;
- maintainers who prefer inspectable automation over opaque end-to-end agents.
The main maintenance work is ordinary research software work:
- add tests for source connectors, ranking, verification, and delivery modules;
- review GitHub Actions workflows for reliability and reproducibility;
- improve configuration validation so new users can diagnose setup errors;
- keep examples and documentation aligned with the code;
- add research-topic presets beyond biomedical imaging;
- prepare releases with notes and known limitations;
- audit report outputs so factual fields are supported by retrieved metadata.
The toolkit is used alongside public PET/MRI and MRI-to-PET research repositories, including:
replicaLT: plasma-guided 3D MRI-to-PET generation;MRI2PET: MRI-to-PET research experiments;ADNI_dataprocess: data processing utilities for ADNI-style neuroimaging workflows.
These repositories provide downstream scenarios for weekly literature monitoring, model-update tracking, experiment reporting, documentation maintenance, and reproducible biomedical AI research workflows.
flowchart LR
A[GitHub Actions schedule] --> B[Search sources]
B --> C[Check title, date, and link]
C --> D[Deduplicate]
D --> E[Rank by relevance and reproducibility]
E --> F[Generate Chinese weekly report]
F --> G[Write Markdown and JSON artifacts]
F --> H[Optional delivery]
Default search flow:
- Search the latest 7 days.
- If strong results are insufficient, extend to 30 days.
- Check metadata before inclusion.
- Keep at most 5 strong results per module and 3 indirect results.
- Produce Markdown and JSON artifacts for every run.
For implementation details, see docs/architecture.md.
Default topics include MRI-to-PET synthesis, pseudo-PET generation, Tau PET, amyloid PET, FDG PET, PET reconstruction, multimodal neuroimaging, and deep learning methods involving PET and MRI.
Default topics include medical vision-language models, biomedical CLIP-style models, radiology foundation models, GitHub repositories and releases, and Hugging Face model or dataset cards.
The generated Chinese weekly report uses six sections:
- 本周期最重要结论
- MRI-to-PET / Tau PET / Alzheimer's disease 强相关论文
- 医学图像大模型 / 医学视觉语言模型更新
- 间接相关但可能有启发的论文或模型
- 未纳入内容与原因
- 下周建议关注关键词
A sanitized demo report is available at docs/demo-email.zh-CN.md.
git clone https://github.com/linshuijin6/research-agent-toolkit.git
cd research-agent-toolkit
python -m pip install --upgrade pip
pip install -e ".[dev]"
cp config.example.yaml config.yaml
rat validate-config --config config.yaml
rat literature-monitor --config config.yaml --dry-runGenerated files are written to outputs/YYYY-MM-DD/.
Typical outputs are email_zh.md, report.json, candidates.json, and excluded.json.
For a more detailed Chinese setup guide, see docs/quickstart.zh-CN.md.
Copy config.example.yaml to config.yaml, then adjust topics, source settings, model endpoint settings, and optional delivery settings for your environment.
By default, the workflow runs in dry-run mode and only writes local artifacts.
The default workflow runs every Monday at 00:00 UTC, which is 08:00 Beijing time.
on:
schedule:
- cron: "0 0 * * 1"
workflow_dispatch:The workflow uploads artifacts and does not send reports unless delivery is explicitly enabled.
Each candidate receives a 0-100 priority score:
[ S = 20\left(0.40R + 0.20N + 0.15C + 0.10P + 0.10Q + 0.05T\right) ]
LaTeX source:
S = 20\left(0.40R + 0.20N + 0.15C + 0.10P + 0.10Q + 0.05T\right)Where R is relevance, N is novelty, C is clinical or research value, P is reproducibility, Q is source quality, and T is timeliness.
- Dry-run is enabled by default.
- Source verification is required by default.
- v1.0 does not read or write Notion.
- The workflow only uses retrieved candidate metadata for report generation.
- Factual fields such as paper titles, DOI values, code links, licenses, model weights, and datasets should be supported by retrieved metadata.
- v1.1: draft review mode improvements.
- v1.2: MCP (Model Context Protocol) adapter.
- v1.3: Notion daily summary workflow.
- v1.4: Web dashboard.
- v1.5: More research-topic presets beyond biomedical imaging.
See the public roadmap for release-readiness tasks and planned extension work.
If this project helps your research workflow, please cite the repository using CITATION.cff.
Apache License 2.0. See LICENSE.