是什么
把"一个模糊的研究问题"拆成多智能体并行检索 + 来源分级 + 交叉验证的七阶段流水线,帮你在几小时内把碎片信息汇成一份带可追溯引用的研究报告,而不是攒一堆链接自己头大。
怎么用
- 先把研究问题、成功标准、输出格式说清楚(这一步省了,后面整份报告都会跑偏)。
- 把大问题拆成 3-5 个子问题,每个子问题指定优先来源类型(学术 / 行业 / 一手数据)。
- 让多个子智能体并行去查,每条结论必须挂上 A-E 来源分级(A 同行评议 / E 道听途说)。
- 进入交叉验证阶段,把不同来源的冲突结论摆出来调和,B 级以下的结论必须标注不确定性。
- 最后做合成 + 引用核查 + 报告封装,交付带执行摘要、参考文献、不确定性说明的成品。
架构图
flowchart LR
A[研究问题] --> B[子问题拆解]
B --> C[多智能体并行检索]
C --> D[来源 A-E 分级]
D --> E[交叉验证]
E --> F[合成报告 + 引用核查]
Deep Research -- Graph of Thoughts Pipeline
Gotchas
Don't skip Phase 1 (scoping). Most research failures come from a vague question, not bad searching. Clarify output format, success criteria, and constraints BEFORE spawning agents.
Sub-agent count matters. 3-5 Web research agents + 1-2 academic agents + 1 cross-validation agent. More than 8 total agents = context explosion with diminishing returns.
Source quality rating is non-negotiable. Every claim needs a grade:
| Grade | What counts |
|---|---|
| A | Peer-reviewed RCT, systematic review, meta-analysis |
| B | Cohort study, clinical guideline, official report |
| C | Expert opinion, case report |
| D | Preprint, conference abstract |
| E | Anecdote, speculation |
Claims below B-grade need explicit uncertainty labels. Don't present D/E sources as facts.
Cross-validation is a phase, not a suggestion. Phase 4 (triangulation) must run before synthesis. Skip it and you get confident-sounding hallucinations.
7-Stage Pipeline
Phase 1: Question Scoping -> clarify with user, define success criteria
Phase 2: Retrieval Planning -> decompose into sub-queries, select sources
Phase 3: Iterative Querying -> spawn sub-agents, execute searches
Phase 4: Source Triangulation -> cross-reference, resolve conflicts, grade sources
Phase 5: Knowledge Synthesis -> structure findings, inline citations
Phase 6: Quality Assurance -> verify citations match content, check for hallucination
Phase 7: Output & Packaging -> format report, executive summary, bibliography
Usage
ait deep-research "research topic"
# Or in conversation: "Deep research [topic]"
Quality Gate
- Every factual claim has a graded, verifiable source
- Key findings confirmed by 2+ independent sources
- Contradictions acknowledged and explained
- No unsupported claims (check for hallucination in Phase 6)
Agent Foundry Team