Evidence-grounded research 证据型深度研究

DeepResearcher Turn complex questions into auditable conclusions. 把复杂问题变成可复核的研究结论。

DeepResearcher scopes the investigation, combines your materials with web and scholarly sources, preserves decision-changing evidence, checks citations and source quality, handles contradictions, and delivers a report with confidence, limitations, and references. DeepResearcher 会界定研究范围，把你的材料与网页、学术来源结合，保留影响决策的证据，核验引用与来源质量，处理矛盾，并交付带置信度、限制说明和参考来源的报告。

Download Orkas and start research 下载 Orkas，开始深度研究

Primary and scholarly sources 优先使用一手与学术来源
Citation verification 引用核验
Contradictions stay visible 矛盾证据不隐藏

DR Research report 研究报告

Research the opportunities and risks of AI note-taking tools in enterprise knowledge management, with cited sources. 系统研究 AI 笔记工具在企业知识管理中的机会和风险，并提供可核验引用。

Evidence status 证据状态 6 sources

Q1 Adoption and use cases 采用现状与用例 Official and first-party evidence 官方与一手证据 verified

Q2 Data and governance risk 数据与治理风险 Standards + peer-reviewed sources 标准与同行评议来源 verified

Q3 Measured productivity impact 生产率影响 Evidence remains mixed 现有证据存在分歧 limited

Sources Citations Coverage Limits

Bounded research plan 有边界的研究计划 Breaks the question into decision-relevant subquestions and a finite source budget. 拆成影响决策的子问题，并控制来源范围。

Evidence ledger 证据记录 Keeps source dates, quotes, limitations, contradictions, and confidence labels traceable. 记录来源日期、引用、限制、矛盾和置信度。

Clean citation gate 引用核验门槛 Unsupported, unproven, or misaligned claims are repaired, weakened, removed, or left unverified. 未支持或错配的结论会被修复、弱化、删除或标记未核验。

Auditable report 可复核报告 Delivers the findings, evidence used, limitations, run status, and a decision-useful next step. 交付结论、使用证据、限制、研究状态和有助决策的下一步。

Usage example 使用示例

Research which positioning claims Orkas should lead with—and which it should not. 研究 Orkas 应该主打哪些产品定位，又不该说哪些。

YOU Research question 研究问题

“Research how product teams evaluate local-first multi-Agent workspaces. Use our twelve interviews and current primary sources to recommend which claims Orkas should lead with on its official-Agent pages.” “研究产品团队如何评估本地优先的多 Agent 工作区。结合 12 份访谈和当前一手来源，建议 Orkas 官方 Agent 页应主打哪些说法。”

12-interviews.pdf product-boundaries.md as-of date decision criteria

Decision 决策目标: Official-page positioning 官方页定位建议
Scope 研究范围: 5 buyer evaluation dimensions 5 个购买决策维度
Cutoff 截止日期: 2026-08-08
Evidence 证据规则: Primary first; conflicting claims retained 一手来源优先；保留矛盾证据

research-report.md 12 interviews + 8 sources · checked 12 份访谈 + 8 个来源 · 已核验

Decision brief · 8 Aug 2026 决策简报 · 2026 年 8 月 8 日

Local-first multi-Agent workspaces: adoption, trust, and buyer fit 本地优先多 Agent 工作区：采用、信任与购买适配度

5 dimensions 12 interviews 8 current sources overall confidence: medium-high 整体置信度：中高

Executive summary 执行摘要

Lead with one-goal orchestration, inspectable specialist handoffs, and local-first workspace materials. Treat provider choice as supporting proof, not the headline. Do not publish unsupported time-saved or headcount-reduction claims. 主打“一个目标的统一调度”、“可检查的专才交接”与“工作区材料本地优先”。模型供应商选择应作为支撑证据，不是主标题。不发布无证据的节时或减员说法。

Users understand the product faster when one outcome stays visible across specialists. 当一个目标在多个专才之间持续可见时，用户更容易理解产品。

Nine of twelve interviews described fragmented context as a bigger problem than raw model capability. 12 份访谈中有 9 份将“上下文分散”视为比单一模型能力更大的问题。 [I2][I5][I9]

Medium 中

Local-first is credible only when its boundary is stated precisely. 只有明确说清边界，“本地优先”才具有可信度。

Workspace files execute locally, while selected models and connectors may still involve their configured providers. 工作区文件在本地执行，但选定模型与连接器仍可能与已配置供应商交互。 [S1][S4]

High 高

Provider choice matters, but setup cost belongs beside it. 模型供应商选择很重要，但必须同时说明配置成本。

Interviewees valued choice after reliability and clear setup guidance; cross-provider cost comparison remains incomplete. 受访者在可靠性与清晰配置指引之后才看重选择权；跨供应商成本对比仍不完整。 [I3][I10]

Medium 中

Evidence matrix 证据矩阵

Claim 结论 Evidence used 使用证据 Confidence 置信度

Workflow clarity 工作流清晰度 12 interviews + shipped orchestration spec 12 份访谈 + 已发布调度规格 Medium 中

Local-first boundary本地优先边界 Product, security, and provider-setting docs 产品、安全与供应商设置文档 High 高

Time saved 节省时间 No controlled before/after measurement 无受控的前后对比测量 Not verified 未核验

Limitations: 限制： The interview sample covers product teams, not enterprise procurement; no controlled productivity study was available at the cutoff. 访谈样本主要是产品团队，不代表企业采购部门；截止日期前也无受控生产率研究。

report.md evidence-ledger.csv source-list.md limitations.md

Research jobs 适合的研究任务

Investigate, compare, verify. 调研、比较、核验。

Industry and landscape research 行业与格局研究

Map candidates, opportunities, risks, and decision criteria from bounded evidence. 基于有限证据梳理候选对象、机会、风险与决策标准。

market landscape
trend report
competitor evidence

Literature and policy review 文献与政策综述

Combine supplied files with peer-reviewed, official, and standards-based sources. 把用户文件与同行评议、官方和标准类来源结合。

literature review
policy comparison
dated source cutoffs

Claim and citation verification 说法与引用核验

Check whether a source exists, matches the claim, and supports the wording actually used. 检查来源是否存在、是否匹配结论、是否支持最终采用的表述。

source quality
quote matching
abstention when needed

How it works 工作方式

Scope, gather, verify, synthesize. 界定、收集、核验、综合。

Scope the decision 界定决策问题

Clarify only the choices that change scope, risk, cost, audience, or deliverable. 只确认会改变范围、风险、成本、受众或交付的关键问题。

Collect changing evidence 收集关键证据

Prefer primary, official, peer-reviewed, dataset, standard, and user-provided sources. 优先使用一手、官方、同行评议、数据集、标准和用户材料。

Verify narrow claims 核验具体结论

Map facts to evidence and remove claims that cannot pass a clean check. 把事实映射到证据，移除无法通过核验的结论。

Deliver the handoff 交付报告

Publish one complete, compact report with findings, limits, sources, and run status. 交付一份完整、紧凑、包含结论、限制、来源和研究状态的报告。

Evidence boundaries 证据边界

A useful partial answer is better than fabricated certainty. 诚实的部分结论，胜过虚构的确定性。

Time-sensitive claims need dates 时效性结论必须带日期

Publication, page update, access, and real-world effective dates are treated separately. 发布日期、页面更新、访问日期和现实生效日期会分别核验。

Source failures stay visible 来源失败不会被隐藏

A failed or inaccessible source becomes a coverage limitation, not an invented replacement fact. 来源无法访问时会形成覆盖限制，不会用臆测事实替代。

High-stakes conclusions remain gated 高风险结论保留专业审查

Research can organize evidence, but legal, medical, financial, and similar decisions still require qualified review. 研究可以整理证据，但法律、医疗、财务等决策仍需专业人士审查。

Works with the team 团队协作

Research can become the evidence layer for the next deliverable. 研究结果可以继续成为后续交付的证据层。

Ask Commander for the complete outcome 把完整目标交给指挥官

Commander can route evidence to ContentWriter for a sourced article, PptMaker for an executive deck, or SeoGeoAgent for answer-ready page content. 指挥官可把证据交给 ContentWriter 写成文章、交给 PptMaker 制作管理层演示，或交给 SeoGeoAgent 形成答案前置页面内容。

Commander ContentWriter PptMaker

Start directly when research is the deliverable 研究本身就是交付时可直接开始

Provide a clear question and optional PDFs, documents, spreadsheets, datasets, or images. DeepResearcher owns the cited report. 给出明确问题，并按需附上 PDF、文档、表格、数据集或图片，DeepResearcher 负责交付引用报告。

Research the next complex question with evidence. 用证据研究下一个复杂问题。

Download Orkas and give DeepResearcher a question, source pack, or set of claims to verify. 下载 Orkas，把研究问题、资料包或待核验说法交给 DeepResearcher。

Download Orkas — free 免费下载 Orkas

Local-first materials · Verifiable sources · Managed or your own models 材料本地优先 · 来源可核验 · 官方模型或自有模型