| name | literature-survey |
| description | Literature survey and review workflow. Triggers on "survey", "综述", "这个方向有什么", "literature review", "related work". |
Literature Survey
Flow
-
Context carry-over:
- Explicit reference ("根据已有的", "用刚才的", "基于这些写综述"): use those papers directly. Skip to step 3 — no candidate listing, no selection question.
- Ambiguous (same topic in context): ASK "刚才找到了 N 篇相关论文,直接用这些?还是再补充搜索?"
- No relevant context: Call
paper_quick_scan(topic, limit=20), then show as table and [FORK] "全部纳入还是选几篇?"
-
Extract structured profiles: Call paper_extract(paper_id) for each selected paper to get structured data (task, method, datasets, baselines, metrics, best_results).
-
Build comparison table: Call paper_compare_table(paper_ids) for data-driven comparison instead of LLM-only generation.
-
Generate survey with structured tables:
- 方法分类表: | 类别 | 代表论文 | 核心思路 | 优势 | 局限 | — built from extracted profiles
- 实验对比表: | 论文 | 数据集 | 指标1 | 指标2 | 最佳结果 | — built from
paper_compare_table data
- 可信度对比: | 论文 | venue | code | 引用数 | 复现风险 | — call
paper_credibility_batch(paper_ids)
- 研究空白与趋势: call
paper_field_stats("method_family") and paper_field_stats("datasets") for data-driven insights
- 结论与建议: 当前方向的成熟度判断、主流方法对比结论、研究机会在哪里
-
Auto-save: call paper_save_report(report_type="survey", content=<survey markdown>, filename="{topic}.md") to persist the survey.
-
[CONTEXT-AWARE FORK] — Based on the survey result, suggest next steps:
-
If survey found clear research gaps:
"文献综述已保存至 {path}(涵盖 {N} 篇论文)。\n💡 下一步建议:\n1. 基于发现的研究空白生成 research ideas\n2. 导出 {N} 篇论文的 BibTeX — 可以直接用在 related work\n(说编号或告诉我你想做什么)"
-
If survey shows a dominant method family:
"文献综述已保存至 {path}。\n💡 下一步建议:\n1. 深入分析 [{dominant_method}] 方向的代表论文 [{top_paper_title}]\n2. 导出 BibTeX 写 related work\n(说编号或告诉我你想做什么)"
-
If survey covers many diverse methods:
"文献综述已保存至 {path}。\n💡 下一步建议:\n1. 看看 [{topic}] 的趋势分析 — 哪个子方向在上升\n2. 导出 BibTeX\n(说编号或告诉我你想做什么)"
If user wants to export
paper_export(paper_ids, format) for BibTeX/markdown
paper_group_add(name="survey-{topic}", paper_ids, create_if_missing=True) to group papers
Quick mode (default)
paper_quick_scan returns 15-20 candidates with one-line summaries. Present this list and let user decide next steps. This IS the survey for most cases.
Full mode
Only when user explicitly asks for "完整综述" / "full survey" / "详细一点":
- Expand to 40+ candidates via
paper_quick_scan(topic, limit=40)
- Generate multi-section survey with per-paper analysis
Rules
- 2 checkpoints max (selection + revision/export). Don't ask about search parameters.
- Default to quick mode. Don't generate a full survey unless asked.
- Always auto-save the survey via
paper_save_report. Additional export (BibTeX) is opt-in via FORK.
- FORK suggestions must reference actual paper counts, method families, and research gaps from the survey result.