Paperclip map/reduce 多论文批读与归纳实战：从字段抽取到结果汇总的完整指南

Paperclip map/reduce 多论文批读与归纳实战从字段抽取到结果汇总的完整指南【免费下载链接】scientific-agent-skillsTurn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 190,000 scientists worldwide. 165 ready-to-use validated skills plus 100 scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.项目地址: https://gitcode.com/GitHub_Trending/cl/scientific-agent-skills导读本指南以 Paperclip 技能中 map-reduce.md 为核心系统讲解如何在不逐篇精读的情况下用map对结果集中的每篇文献并行运行 LLM 阅读器抽取字段再用reduce把各篇答案综合成单一输出从而回答横跨多篇论文的问题。读完本文你将掌握map三种 worker 的适用场景、查询词写法、结构化输出与断点续跑、reduce六种策略及两个已知缺陷的规避方法以及配套的results导出和ask-image图表问答能力。Paperclip 的 map/reduce 定位跨论文问答的核心流水线Paperclip CLI 把约 1100 万篇全文论文、21.7 万监管文档、11 万临床试验方案和 57.4 万蛋白质条目暴露为一个只读虚拟文件系统配合服务端语义检索和 LLM 阅读器使用见 SKILL.md。在这个体系里map与reduce承担跨文档综合这一层职责map对一个结果集result set中的每篇文档并行运行一个 LLM 阅读器逐篇产出答案reduce把这些逐篇答案综合成一个输出。二者组合起来就是回答一个横跨多篇论文的问题而不必亲自逐篇通读的标准姿势。需要特别强调的是reduce只综合map已返回的内容不会重新阅读论文——如果reduce输出里缺了某个字段说明它本来就缺失在map结果里正确做法是修正 map 查询后重跑而不是去 reduce 里找补。整条流水线的典型形态如下示例省略了认证前缀真实每次调用都需带上详见下文认证前缀一节paperclip search -s pmc lipid nanoparticle mRNA delivery -n 10 # → s_abc123 paperclip filter --from s_abc123 in vivo delivery with quantified efficiency paperclip map --from s_abc123 What delivery vector, target cell type, and transfection efficiency were reported? paperclip reduce --from m_def456 --strategy table Compare vector, cell type, and efficiency结果 id 的命名约定整条链路上每个阶段都会打印一个结果 id后续命令靠它衔接前缀来源说明s_*search/filter/grep检索/过滤/全文 grep 产生的结果集m_*map逐篇阅读输出r_*reduce归约产物如Artifact ID: r_36626b45reduce在省略--from时默认取最近一次 map但文档明确建议总是显式传--from——这是防止误捡到无关运行的低成本保险。map 完成后的终端输出与.gxl陷阱一次完成的 map 会打印进度条、每篇论文的预览每篇约截断成一行以及一个Full results: /.gxl/map_id.txt指针[######..............] 1/3 papers run m_4b4632df [2s] Map complete: 3/3 tasks succeeded in 3571ms Results ID: m_4b4632df Full results: /.gxl/map_m_4b4632df.txt [success] Piperazine-derived lipid nanoparticles deliver mRNA to immune cells in (PMC9376583) Based on the paper, here are the details ... * **Delivery Vector:** Piperazine-derived lipi关键缺陷已验证这个/.gxl/路径无法被读取——cat它返回 No such file尽管ls /.gxl/能看到该文件。cli-reference.md 进一步印证map、reduce都会把转录写入/.gxl/如reduce_r_36626b45.txt、map_m_4b4632df.txt但cat一律报错。要看未截断的逐篇答案必须改用resultspaperclip results m_4b4632df # full output, with real document ids paperclip results m_4b4632df --save map.csv # or export itmap对结果集逐篇运行 LLM 阅读器完整参数表以下参数来自paperclip map --help0.7.14–0.7.15 版本验证map --from RESULTS_ID [OPTIONS] query --from ID Result id from a previous search (required) --worker NAME quick-reader (default) | eligibility-screen | exhaustive-extraction --output_schema JSON Structured output schema --claim-schema JSON JSON Schema each exhaustive claim must satisfy --repo NAME Shared repo receiving validated exhaustive claims --resume MAP_ID Continue pending work; never reruns successful papers --retry-failed With --resume, also retry failed papers --cancel MAP_ID Durably request cancellation; pending work will not start -n, --limit N Limit number of papers processed --offset N Skip the first N papers -j, --max-concurrent N Concurrent extraction subagents (default 100, server hard cap 256)三种 worker 的选型Worker适用场景quick-reader默认普通字段抽取和逐篇问答每篇一次 LLM 调用eligibility-screen对整篇论文做单轮结构化筛选依据纳入/排除标准——即系统综述的筛选步骤exhaustive-extraction多轮 Claude 工具型 worker逐项检查 Methods、Results、Tables、Figures 和补充材料。慢但彻底用于定量字段抽取查询词写法map 成败的关键文档直言这是 map 成败的分水岭并给出四条经验把你想要的每个字段都枚举出来——worker 只会返回你明确要求的东西不会多做知道章节名就指名道姓例如 From the Methods section, extract the cell line, passage number, and transfection reagent.显式要求缺失情况的表达If the paper does not report a sample size, say not reported.——否则你无法区分论文没报和抽取漏了反例与正例对比差Summarize this paper.好What delivery vector was used, what cell type was targeted, and what transfection efficiency was reported? State not reported for any field the paper omits.另外search-and-retrieval.md 指出搜索阶段就应控制规模Use-n 5/-n 10beforemap因为 map 对每篇论文都是一次付费 LLM 调用先filter再map更省钱。规模控制Sizing交互式工作建议把quick-reader控制在3–10 篇每篇一次 LLM 调用更大规模时宁可一次 map 更高的-j也不要发多个重叠的 map 请求服务端把并发硬上限设在256单用户配额可能更低map 完成后直接基于其输出作答不要回头逐篇重读论文——那等于把你刚付过钱的工作丢掉。结构化输出--output_schema与--claim-schemapaperclip map --worker eligibility-screen \ --output_schema {decision:yes|no|uncertain,reason:string} \ --from s_abc123 Apply the protocol eligibility criteriapaperclip map --worker exhaustive-extraction \ --repo review \ --claim-schema {type:object,required:[type],properties:{type:{type:string}}} \ --from s_yes Extract all requested claims--output_schema约束单条输出为指定 JSON 结构适合系统综述的资格筛选yes|no|uncertain 理由--claim-schema--repo把每条通过校验的 exhaustive claim 直接路由进一个共享 repo——这正是paperclip-meta-analysis工作流的底层机制相关 repo 命令如repo add、repo commit、repo status见 cli-reference.md。文档提醒只有用户明确要求可验证语料库时才使用它。长任务--resume/--retry-failed/--cancelpaperclip map --resume m_abc123 # continue; successful papers are not redone paperclip map --resume m_abc123 --retry-failed # also retry failures paperclip map --cancel m_abc123 # stop pending workresume 是持久化的大型抽取若撞上超时可恢复而不必重头再来——优先--resume --retry-failed而不是整体重启。--cancel发出的是持久化的取消请求未开始的待办任务将不再启动。reduce把逐篇答案综合成单一输出完整参数与六种策略reduce --from MAP_ID [OPTIONS] question --from ID Map result id (m_*); defaults to the most recent map --strategy STR summarize (default) | table | themes | consensus | bullet_points | extract --columns COL,... Comma-separated columns for the table strategy策略产出summarize跨论文的整合叙事默认table返回散文而非表格——见下方缺陷说明themes反复出现的主题与分组consensus论文一致与分歧之处——有争议结论时的正确选择bullet_points浓缩要点列表extract只给抽取到的值极少散文已验证缺陷一--strategy table不出表格在 0.7.14 和 0.7.15 上都验证过无论是否带--columns输出都是多段散文。如果需要对比表格请用paperclip results m_id拿全量数据自行构建。这个缺陷在 SKILL.md 的Known defects表和 cli-reference.md 中均有记录paperclip reduce --from m_def456 --strategy table \ --columns paper,vector,cell type,efficiency,n \ Compare delivery approaches paperclip reduce --from m_def456 --strategy consensus \ Do these studies agree on whether LNP delivery reaches hematopoietic stem cells in vivo?已验证缺陷二reduce 内嵌的引用标记 id 被截断reduce 的散文里会携带内联引文标记形如... cholesterol, DMG-PEG2000, and DOPE or DSPC {{document_id: PMC12388, line: 5}}这些文档 id 被截断到 8 个字符无法解析。上面这篇真实论文是PMC12388858而PMC12388会返回cat: PMC paper not found。同样的还有PMC93765实为PMC9376583和PMC11843实为PMC11843327。因此用 reduce 标记拼出的引用 URL 必是死链SKILL.md 称这是最严重的一个缺陷。正确姿势只把标记当作去哪里找的线索真实 id 一律从search、results或meta.json取paperclip results m_4b4632df # real ids paperclip head -50 /papers/PMC12388858/content.linesreduce 输出本身不可引用。在答案里引用某个数字前必须打开它来源的论文、读到那一行并引用该行——map 和 reduce 都是 LLM 摘要器而引用契约要求你引用的文本是你确实亲眼读过的内容。results读取、导出与找回 idpaperclip results --list # recent ids with the command that made each one paperclip results s_4a2b61f6 # view a saved result set paperclip results s_4a2b61f6 --save out.csv # export to CSV paperclip results m_def456 --save map.txt # export to TXT--list输出形态Recent results (20): s_3b1a8db3 search -s papers somatic hypermutation -n 2 2026-07-28 01:00 s_a5590fe3 grep -l SLC30A8 /papers/ 2026-07-28 00:58丢 id 时或想拿新检索和旧检索对比时非常有用。另需注意map 的终端预览被截断、/.gxl/不可读因此results是获取逐篇完整答案的唯一可靠途径。ask-image对论文图表提问ask-image PATH question ask-image --list # figures in the current directory (requires cd into a paper) --fn describe Describe the figure --fn extract-data Extract data from the figure务必先ls。图表文件名来自出版商而非fig1.jpg这种约定命名猜名字会以Error: Image not found失败paperclip ls /papers/PMC10945750/figures/ # pnas.2307796121fig01.gif pnas.2307796121fig01.jpg paperclip ask-image /papers/PMC10945750/figures/pnas.2307796121fig01.jpg \ What is on each axis, which conditions are compared, and what is the reported effect size? paperclip ask-image /papers/PMC10945750/figures/pnas.2307796121fig01.jpg --fn extract-data--list虽被文档列为备选但它要求先cd进论文目录而cd在多次调用间不持久——所以直接用ls更可靠。对图表的视觉抽取只是估算。如果某个数字很关键应从正文或补充材料里找到它再引用paperclip ls /papers/PMC10945750/supplements/ paperclip head -40 /papers/PMC10945750/supplements/file成本与失败处理速查quick-reader每篇一次 LLM 调用exhaustive-extraction是多轮、重得多的调用。检索结果超过约 10 条时先filter再map——先丢弃无关论文更划算。map可以直接作用于grep的结果 idgrep 和 search 一样返回s_*id。运行中途有论文失败时用--resume --retry-failed而非重来。空白的逐篇答案通常意味着查询提到了论文没有报告的字段而不是阅读器失败。显式索要 not reported 才能区分这两种情况。前置提醒认证前缀与配套阅读本指南所有示例省略了认证前缀。每次真实调用都必须加上它shell 状态在工具调用间不保留裸跑会静默回退到另一个身份[ -f .env ] { set -a; . ./.env; set a; }; paperclip command若想深入上下游能力建议继续阅读同仓库配套文档search-and-retrieval.md检索、filter、grep 与结果读取、cli-reference.md全部命令与参数、虚拟文件系统与沙箱限制、repos-and-workspace.mdrepo、claim 与上传导出以及 installation.md安装、认证与各客户端 MCP 配置。总结mapreduce把跨论文综合从人工苦力变成一条可审计的流水线search/grep得结果集 →filter精简 →map逐篇抽取 →reduce综合 →results导出。用好它的关键在四点查询词显式枚举字段并索要 not reported、控制 map 规模并显式传--from、用results而非/.gxl/取全量输出、以及牢记两个已验证缺陷——--strategy table实际返回散文、reduce 的引文 id 被截断不可用。守住reduce 输出不可直接引用、引用前必须亲眼读行这条底线你就能在多文献场景中既高效又可靠。【免费下载链接】scientific-agent-skillsTurn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 190,000 scientists worldwide. 165 ready-to-use validated skills plus 100 scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.项目地址: https://gitcode.com/GitHub_Trending/cl/scientific-agent-skills创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考