Research

Formulate 2–4 search queries that cover:

1. The core topic (broad)
2. Specific subtopics or recent developments
3. Comparisons, applications, or critiques (if relevant)

3. Search the Web

Execute web searches:

web_search(query="<query 1>", count=5)

web_search(query="<query 2>", count=5)

Collect all result URLs and titles. Deduplicate URLs across searches.

4. Search Academic Papers (If Academic)

For academic or scientific topics, also search arXiv:

arxiv(query="<academic query>", max_results=5)

• Note the paper titles, arXiv IDs, and abstracts from results.
• Select the 2–3 most relevant papers for deeper reading.

For important papers, download and read them:

exec(command="mkdir -p research_papers && curl -L -o research_papers/<arxiv_id>.pdf https://arxiv.org/pdf/<arxiv_id>.pdf", timeout=120)

On Windows:

exec(command="if not exist research_papers mkdir research_papers && curl -L -o research_papers/<arxiv_id>.pdf https://arxiv.org/pdf/<arxiv_id>.pdf", timeout=120)

Then extract key content:

pdf_reader(path="research_papers/<arxiv_id>.pdf", start_page=1, end_page=5)

Focus on Abstract, Introduction, and Conclusion sections to extract core findings efficiently.

5. Fetch Web Page Content

From the search results in Step 3, select the top 3–5 most relevant URLs and fetch their content:

web_fetch(url="<url>", extractMode="markdown")

For each fetched page:

• Extract key facts, data, quotes, and arguments.
• Note the source URL for citation.
• Skip pages that are paywalled, empty, or irrelevant.

If a fetch fails, try the next URL from your list. Do not waste iterations on broken links.

6. Analyze and Synthesize

Before writing the report, organize your findings:

1. Identify key themes: Group information by subtopic or theme.
2. Cross-reference sources: Note where multiple sources agree or disagree.
3. Assess reliability: Prefer peer-reviewed papers and authoritative sources over blog posts or opinion pieces.
4. Identify gaps: Note areas where information is insufficient or conflicting.

7. Write the Research Report

Compose a comprehensive Markdown report. The report MUST follow this structure:

# [Research Topic Title]

> **研究日期**: [current date, YYYY-MM-DD]
> **研究类型**: [学术研究 / 行业研究 / 综合研究]
> **数据来源**: Web搜索, [arXiv论文 if applicable], 网页内容抓取

---

## 摘要

[A concise 150–300 word executive summary of the key findings.]

## 1. 背景介绍

[Context and background information on the topic. Why is this topic important? What is the current state of the field?]

## 2. 主要发现

### 2.1 [Subtopic / Theme 1]

[Detailed findings with citations.]

### 2.2 [Subtopic / Theme 2]

[Detailed findings with citations.]

### 2.3 [Subtopic / Theme 3]

[Detailed findings with citations.]

[Add more subsections as needed.]

## 3. 学术研究综述

> Include this section ONLY for academic topics.

[Summary of key academic papers found, their methods, and conclusions. Reference specific papers by title and authors.]

| 论文 | 作者 | 年份 | 核心贡献 |
|------|------|------|----------|
| [Paper Title 1] | [Authors] | [Year] | [Key contribution] |
| [Paper Title 2] | [Authors] | [Year] | [Key contribution] |

## 4. 分析与讨论

[Your synthesis of the findings. Identify trends, contradictions, and implications. Compare different viewpoints.]

## 5. 结论与建议

[Concise conclusions based on the research. Actionable recommendations if applicable.]

## 6. 研究局限性

[Note any limitations: topics not fully covered, potential biases in sources, paywalled content that could not be accessed, etc.]

## 参考来源

[List all sources used, formatted as:]

1. [Title](URL) — [brief description of what was extracted]
2. [Title](URL) — [brief description]
3. [Paper Title, Authors, Year](arXiv link) — [brief description]

Adapt the section titles to the user's language. If the user writes in English, use English headings; if in Chinese, use Chinese headings as shown above.

8. Save the Report

Create the output directory and write the report:

write_file(path="research/<topic_slug>.md", content="<full report markdown>")

Use a URL-safe slug for the filename (e.g. research/llm-fine-tuning-2025.md).

Confirm the file was saved:

list_dir(path="research")

Report to the user:

• Output file path
• Number of sources consulted
• Key findings summary (2–3 sentences)

Tips

• Efficiency: Don't fetch every search result. Prioritize the top 3–5 most relevant URLs based on title and snippet.
• Token management: When fetching long pages, use maxChars parameter to limit content: web_fetch(url="...", maxChars=20000).
• Academic depth: For academic topics, reading paper abstracts and conclusions is usually sufficient. Only read full papers if the user specifically requests deep analysis.
• Iterative search: If initial searches don't yield good results, refine your queries with more specific terms or synonyms.
• Freshness: Prefer recent sources. Note publication dates when available.
• The research/ directory is relative to the workspace root. Create it if it doesn't exist.
• The research_papers/ directory stores downloaded PDFs for academic research. Create it only when needed.
• Multiple queries: Use varied queries to avoid echo-chamber results. Include different keywords, phrasings, and angles.
• Fallback: If web_search returns poor results, try rephrasing the query or using more specific terminology. If arxiv returns no results, the topic may not have academic coverage — note this in the report.