# Extension Path: For Researchers > [繁體中文](./for-researcher.md) | [简体中文](./for-researcher.zh-Hans.md) | **English** [← Back to the main route](../README.en.md) ## 📌 What this path helps you do This page does not make AI your researcher. It helps with one simpler task: **find sources, understand them, and confirm that answers are actually supported by evidence.** - If you use a terminal or Python, come after [Track A A3](../tracks/cli/A3-cli-production.en.md) or [Track B Stage 7](../stages/07-multi-agent-production.en.md). - If you do not code, start with the first exercise below. You only need a browser and one public paper. ## 🎯 Learning goals After this page, you can: 1. Separate what AI says from what the original text says. 2. Check each numbered reference instead of trusting an answer just because it has reference numbers. 3. Know which data may be uploaded and which requires permission from an institution or data owner. 4. Keep enough records for yourself or a colleague to reproduce the work. ## 🧩 Eight core terms - **Source**: original material used for verification, such as a paper, dataset, or research record. - **Claim**: a checkable statement, such as “method A performs better on dataset B.” - **Citation**: a signpost back to a source location; it does not guarantee that the source supports the claim. - **Source Verification**: open the original and check that content, scope, and limitations match the answer. - **Literature RAG**: retrieve passages from permitted literature, then give them to a model to answer. - **Reproducibility**: others can rerun comparable results from your data, steps, versions, and settings. - **Private Data**: content that cannot be freely published or uploaded, such as participant data, medical records, unpublished manuscripts, or company secrets. - **Human Review**: a person is responsible for claims, citations, code, tables, and the final decision; AI cannot sign or assume responsibility. ## 🛠 First exercise: verify three answers about one paper Before uploading, confirm that the **license or copyright** and the **tool's terms** allow it. Publicly readable is not permission to upload a paper to another service. Use the public paper [Attention Is All You Need](https://arxiv.org/abs/1706.03762). Add the paper to a citation-capable tool and paste: ```text Answer only from this paper. Attach a citation to each answer; if evidence is missing, write “unsupported” and do not guess. 1. What problem does the paper aim to solve? 2. What are the main parts of the proposed method? 3. Which experiments support the result, and what limitations do the authors state? After answering, list the original text for each citation. Do not present your inference as an author claim. ``` Then open each citation, read answer and original text together, and mark unsupported sentences **unsupported** instead of adding an unrelated citation. ## 📚 Choose an entry point

What you want	Start with	Why	Rating
Ask about one paper in a browser	Gemini Notebook (formerly NotebookLM)	Return from source uploads to citations	⭐⭐⭐⭐⭐
Organize your literature library	Zotero	Organize PDFs, authors, years, and notes first	⭐⭐⭐⭐⭐
Build rerunnable literature RAG in Python	PaperQA2	Science-document and citation-centered workflow	⭐⭐⭐⭐⭐

Gemini Notebook is Google’s current name for NotebookLM as of 2026-07-16; the old name remains for recognition. A citation is an entry point for checking, not a guarantee. ## 📖 Required reading Read in order. The first two prevent treating citations as guarantees; the next four help preserve sources, code, data, and results: 1. [Gemini Notebook citation help](https://support.google.com/gemininotebook/answer/16179559): open citations and read context. 2. [Gemini Notebook privacy and terms](https://support.google.com/gemininotebook/answer/17004255): understand data handling before upload. 3. [Zotero quick start](https://www.zotero.org/support/quick_start_guide): organize authors, years, PDFs, and notes. 4. [PaperQA2 README](https://github.com/Future-House/paper-qa): connect literature RAG answers to documents. 5. [DVC command reference](https://doc.dvc.org/command-reference): version data and rerunnable pipelines with Git. 6. [Zenodo quickstart](https://help.zenodo.org/docs/get-started/quickstart/): preserve publishable data, code, or materials in a citable version. ## ⭐ Curated research tools and projects Tool names, licenses, and repository status were checked against official pages and the GitHub API on 2026-08-29 UTC. Ratings are editorial ratings for this map, not GitHub stars or rankings.

Group	Project / document	Rating	Use	Limits
Start and organize	Gemini Notebook (formerly NotebookLM)	⭐⭐⭐⭐⭐	Source-grounded Q&A and citations	Available; cloud service Check every citation; review policy before private data
	Zotero	⭐⭐⭐⭐⭐	Manage PDFs, metadata, notes, and citations	Available; desktop / web Manages sources; does not judge research quality
	paper-qa	⭐⭐⭐⭐⭐	Build citation-grounded literature RAG in Python	Active; Apache-2.0 Configure model and sources; evaluate quality yourself
Explore and write	gpt-researcher	⭐⭐⭐⭐	Multi-source search and research briefs	Active; Apache-2.0 Find candidate sources; not the final citation judge
	storm	⭐⭐⭐⭐	Organize viewpoints, outlines, and long-form writing	Usable; MIT; slower updates Check dependencies and sources before use
	ChatPaper	⭐⭐⭐⭐⭐	Chinese paper summaries, translation, and writing support	Usable; CC BY-NC-ND 4.0 Noncommercial, no-derivatives license; not a general open-source license
	zotero-gpt	⭐⭐⭐⭐	Interact with literature in Zotero	Usable; AGPL-3.0 Maintain plugin and model settings separately
Repeatable research and evidence	asreview	⭐⭐⭐⭐	Active-learning support for systematic-review screening	Active; Apache-2.0 Ranking saves time; human screening still decides inclusion and keeps the record
	dvc	⭐⭐⭐⭐⭐	Keep data versions, models, and pipelines rerunnable	Active; Apache-2.0 Needs Git and storage; versions do not prove conclusions
	mlflow	⭐⭐⭐⭐⭐	Track parameters, metrics, data, and artifacts across runs	Active; Apache-2.0 Tracking does not make an experiment valid; keep secrets and participant data out
	Zenodo	⭐⭐⭐⭐⭐	Publish data, code, and materials with a DOI	Available; cloud service Metadata is public; de-identify private data under institutional rules
	repo2docker	⭐⭐⭐⭐	Rebuild a runnable environment from repository settings	Active; BSD-3-Clause A container preserves the environment; also preserve data, hardware needs, and external services
Research automation	flonat-research	⭐⭐⭐	Research skills, agents, hooks, and LaTeX workflows	Active; MIT Infrastructure example, not universal for every field
Research automation	AI-Scientist-v2	⭐⭐⭐⭐	End-to-end multi-agent research experiments	Research reference; custom source-code license License requires disclosure of machine-generated manuscripts; authors remain responsible
History	open_deep_research	⭐⭐⭐	Study early deep-research agent architecture	Archived; MIT; historical reference Not a current default

## ✅ Completion check and next stop - [ ] I checked three answers, not just citation numbers. - [ ] I found an example supported by the original or marked unsupported. - [ ] I did not upload unapproved Private Data. - [ ] I saved sources, questions, tool name, date, and my judgment. Next: use [Stage 6](../stages/06-memory-rag.en.md) for literature RAG; [Stage 7](../stages/07-multi-agent-production.en.md) for multiple agents; and the [MCP / Skills catalog](../resources/mcp-skills-catalog.en.md) for external tools.

⏱ Expand: time, accounts, cost, and data safety

The first exercise takes about 20–40 minutes. For Private Data, pause and confirm IRB, institutional policy, contracts, data-owner consent, and tool terms. [Gemini Notebook privacy guidance](https://support.google.com/gemininotebook/answer/17004255) says general content is not directly used to train foundation models unless feedback is provided, and feedback may be reviewed by people; this does not automatically approve research uploads. Plans, quotas, and account rules change, so check official pages rather than preserving fixed prices.

🧪 Expand: turn one-paper practice into a rerunnable workflow

### Literature inbox Save DOI, URL, authors, year, and acquisition date; let tools summarize while linking each claim to the original; humans decide read, exclude, or verify and record why. ### Cross-paper synthesis Ask what each paper says before comparing agreement, conflict, and conditions. Do not ask for a complete story before finding citations. ### Code and experiments Save data versions, environment, seed, prompt, model/tool versions, outputs, and human edits. Rerunning does not prove correctness, but missing records hide errors. ### Before submission Check every claim, citation, table, figure, program, and journal rule. Authors make the final judgment and disclose use under journal policy.

🧯 Expand: common errors, alternatives, and troubleshooting

| Problem | What to do first | |---|---| | Citation does not support the answer | Mark unsupported, narrow the question, and do not force a related citation | | Tool cannot read a scanned PDF | OCR first, then spot-check pages and formulas | | Conclusions from papers are mixed | Require paper name and page/paragraph for each claim before synthesis | | Data cannot go to the cloud | Use an institutional environment; consider the local RAG route in [Stage 6](../stages/06-memory-rag.en.md) | | Automation is too complex | Return to one paper, three questions, and one-by-one checking | No tool replaces IRB, data governance, author responsibility, or domain expertise.