查证 · Deep Research 出一份带引用的报告Verify · A Cited Report from Deep Research.
有些问题你要的不是一段聊天回答，是一份能拿去用、每条结论都带出处的报告。Deep Research 会自己读大量来源、交叉核对、写成结构化文档 —— 代价是它要跑几分钟到几十分钟。这章讲什么活值得等这一趟，什么时候普通对话或一次搜索就够。Some questions do not want a chat reply — they want a usable report where every claim carries a source. Deep Research reads across many sources on its own, cross-checks them, and writes a structured document — at the cost of minutes to tens of minutes. Which work is worth the wait, and when an ordinary chat or a single search is enough.
有些问题你要的不是一段聊天回答，而是一份能拿去开会、写方案、做采购判断的报告：来源要清楚，结论要能追溯，范围要能说明。Some questions do not want a chat reply. They want a report you can take into a meeting, plan, or purchasing decision: sources visible, conclusions traceable, scope explainable.
— I
先审计划，再让它跑Review the Plan Before It Runs.
官方说明现在把流程写得很清楚：你描述结果、选择来源，ChatGPT 生成研究计划；你可以审查和修改计划，再让它开始；运行时还能看进度、中途打断、调整来源；最后得到带 citations 或 source links 的结构化报告。1注 1Note 1OpenAI 帮助中心 · Deep Research —— Deep Research 会规划、研究、综合复杂问题并产出带 citations / source links 的报告；用户可审查和修改研究计划、跟踪进度、中途打断或调整来源。OpenAI Help · Deep Research — Deep Research plans, researches, and synthesizes complex questions into reports with citations / source links; users can review and modify the research plan, follow progress, interrupt, or adjust sources.这意味着 Deep Research 不是黑盒。你不应该把一句模糊问题扔进去等半小时，而要把它的计划当作第一版研究提纲来审。The current official guidance lays out the flow clearly: describe the outcome, choose sources, and ChatGPT creates a research plan; you can review and modify it before it begins; while it runs, you can watch progress, interrupt, and adjust sources; the result is a structured report with citations or source links.1注 1Note 1OpenAI 帮助中心 · Deep Research —— Deep Research 会规划、研究、综合复杂问题并产出带 citations / source links 的报告；用户可审查和修改研究计划、跟踪进度、中途打断或调整来源。OpenAI Help · Deep Research — Deep Research plans, researches, and synthesizes complex questions into reports with citations / source links; users can review and modify the research plan, follow progress, interrupt, or adjust sources. That means Deep Research is not a black box. Do not throw in a vague question and wait half an hour; treat its plan as a first research outline to review.
提示词PromptDeep Research 提问Deep Research prompt
研究问题：<要查清楚的问题>
用途：<我要用它做什么决定>
范围：<时间窗 / 地区 / 行业 / 来源类型>
排除：<不要看的来源或角度>
必须比较：<维度 1>、<维度 2>、<维度 3>
产出：先给研究计划等我确认；最终给对比表、结论、风险和每条关键结论的来源链接。Research question: <question to resolve>
Use: <decision this will support>
Scope: <time window / region / industry / source type>
Exclude: <sources or angles to skip>
Must compare: <axis 1>, <axis 2>, <axis 3>
Output: first give a research plan for my approval; final output should include a comparison table, conclusions, risks, and source links for each key claim.
研究问题：2026 年主流向量数据库的自托管选型，哪个最适合中小团队？
用途：决定我们新项目的向量检索用哪个，要能说服团队。
范围：2024-01 至 2026-06，优先官方文档和 GitHub，含云托管和自托管。
排除：第三方营销软文、没有数据支撑的博客。
必须比较：部署复杂度、查询性能、社区活跃度。
产出：先给研究计划等我确认；最终给对比表、结论、风险和每条关键结论的来源链接。Research question: among mainstream vector databases in 2026, which self-hosted option best fits a small-to-mid team?
Use: decide which vector store our new project uses, with enough to convince the team.
Scope: 2024-01 to 2026-06, prioritizing official docs and GitHub, covering both managed and self-hosted.
Exclude: third-party marketing posts and blogs without supporting data.
Must compare: deployment complexity, query performance, community activity.
Output: first give a research plan for my approval; final output should include a comparison table, conclusions, risks, and source links for each key claim.
计划里有没有明确时间窗、地域、对象和排除项？Does the plan specify time window, region, subjects, and exclusions?
来源是否包括你真正信任的官方页、原始文件或指定站点？Do sources include the official pages, primary documents, or specified sites you trust?
产出形态是否写清：表格、结论、风险、建议？Is the output shape clear: table, conclusions, risks, recommendation?
如果它看不到某个内部来源，你是否准备好上传文件或启用对应 app？If it cannot see an internal source, are you ready to upload files or enable the right app?
— II
用一次额度，买一份报告Spend One Allowance Unit on a Report.
Deep Research 适合多步、深入、需要跨多个来源综合的问题，尤其是你想控制来源范围时。官方也提醒，快速查找或短对话用标准 chat 会更快。1注 1Note 1OpenAI 帮助中心 · Deep Research —— Deep Research 会规划、研究、综合复杂问题并产出带 citations / source links 的报告；用户可审查和修改研究计划、跟踪进度、中途打断或调整来源。OpenAI Help · Deep Research — Deep Research plans, researches, and synthesizes complex questions into reports with citations / source links; users can review and modify the research plan, follow progress, interrupt, or adjust sources.用量也不是无限：不同计划不同，产品内 counter 显示剩余任务；固定月度额度从首次使用日期起每 30 天重置。2注 2Note 2OpenAI 帮助中心 · Deep Research usage —— 可用量随计划变化，产品内 counter 显示剩余任务；固定月度额度从首次使用日期起每 30 天重置。OpenAI Help · Deep Research usage — allowances vary by plan, the in-product counter shows remaining tasks, and fixed monthly allowances reset every 30 days from first use.所以它不是默认搜索框，而是「一份额度换一份报告」的入口。Deep Research fits multi-step, deep questions that synthesize across sources, especially when you want to control source scope. Official guidance also notes that standard chat is faster for quick lookups or short conversations.1注 1Note 1OpenAI 帮助中心 · Deep Research —— Deep Research 会规划、研究、综合复杂问题并产出带 citations / source links 的报告；用户可审查和修改研究计划、跟踪进度、中途打断或调整来源。OpenAI Help · Deep Research — Deep Research plans, researches, and synthesizes complex questions into reports with citations / source links; users can review and modify the research plan, follow progress, interrupt, or adjust sources. Usage is not infinite: it varies by plan, the product counter shows remaining tasks, and fixed monthly allowances reset every 30 days from first use.2注 2Note 2OpenAI 帮助中心 · Deep Research usage —— 可用量随计划变化，产品内 counter 显示剩余任务；固定月度额度从首次使用日期起每 30 天重置。OpenAI Help · Deep Research usage — allowances vary by plan, the in-product counter shows remaining tasks, and fixed monthly allowances reset every 30 days from first use. So this is not your default search box; it is the entrance where one allowance unit buys a report.
任务Task
入口Entrance
查一个单点事实、改一句话、问一个 how-toLook up one fact, rewrite one sentence, ask one how-to
Chat
需要跨多个来源综合、带出处、可交付Synthesize across sources, cite, deliver
Deep Research
需要登录网站、点击、填写、下载、整理Login, click, fill, download, organize
Deep Research 和 agent 会相邻、也会重叠。agent 也能研究，Deep Research 也能用 apps 和指定来源；分界不是工具名，而是目标。要一份带引用、可审查的报告，用 Deep Research。要浏览网页、登录、填表、整理文件，用 agent。3注 3Note 3OpenAI 帮助中心 · ChatGPT agent —— agent 面向浏览网站、处理文件、连接第三方数据源、填表和编辑表格等行动任务；和 Deep Research 的报告任务分工不同。OpenAI Help · ChatGPT agent — agent is for action tasks such as navigating websites, working with files, connecting third-party sources, filling forms, and editing spreadsheets; its job differs from Deep Research's report job.Deep Research and agent sit next to each other and sometimes overlap. Agent can research, and Deep Research can use apps and selected sources; the boundary is not the tool name but the objective. For a cited, reviewable report, use Deep Research. For browsing, login, form-filling, and file handling, use agent.3注 3Note 3OpenAI 帮助中心 · ChatGPT agent —— agent 面向浏览网站、处理文件、连接第三方数据源、填表和编辑表格等行动任务；和 Deep Research 的报告任务分工不同。OpenAI Help · ChatGPT agent — agent is for action tasks such as navigating websites, working with files, connecting third-party sources, filling forms, and editing spreadsheets; its job differs from Deep Research's report job.Deep Research 可使用公共 web、上传文件、指定站点和已启用 apps；Apps 的能力又会随计划和 workspace 设置变化。4注 4Note 4OpenAI 帮助中心 · Apps in ChatGPT —— Apps 可支持 search、deep research、sync、write 和 custom MCP 等能力，具体能力按计划与 workspace 配置变化；Deep Research 可使用已启用 apps。OpenAI Help · Apps in ChatGPT — Apps can support search, deep research, sync, write, and custom MCP capabilities, with availability varying by plan and workspace configuration. Deep Research can use enabled apps.这带来一个现实边界：它未必能看到你以为它能看到的来源。关键事实、价格、法律、医疗、金融、技术版本，仍要人工抽查，特别是会影响购买、发布、签约和上线的结论。Deep Research can use the public web, uploaded files, specified sites, and enabled apps; app capabilities vary by plan and workspace settings.4注 4Note 4OpenAI 帮助中心 · Apps in ChatGPT —— Apps 可支持 search、deep research、sync、write 和 custom MCP 等能力，具体能力按计划与 workspace 配置变化；Deep Research 可使用已启用 apps。OpenAI Help · Apps in ChatGPT — Apps can support search, deep research, sync, write, and custom MCP capabilities, with availability varying by plan and workspace configuration. Deep Research can use enabled apps. That creates a practical boundary: it may not see the sources you think it can see. Key facts, prices, legal, medical, financial, and technical-version claims still need human spot checks, especially when they affect purchasing, publishing, contracting, or shipping.
只抽查会影响决策的结论，不必逐字核完整报告。Spot-check decision-driving claims; do not try to verify every sentence.
点开来源看原文是否真的支持结论。Open the source and confirm it actually supports the claim.
数字、价格、日期、版本、法律条款必须二次确认。Numbers, prices, dates, versions, and legal terms need second confirmation.
如果来源缺失或只来自二手转述，让它补官方或一手来源。If sources are missing or second-hand, ask for official or primary sources.
Deep Research 替你读,
引用还得你抽查.
Deep Research reads for you,
you still spot-check the citations.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return