AI Deep Research Assistant

Vague Idea → Multi-Perspective Research → Academic-Grade Report

Not just search — real research.
Stanford STORM methodology · Multi-perspective analysis · Simulated expert dialogue · Fact verification

---

English | 中文

---

🇬🇧 English

Why This Exists

Traditional Workflow	This Tool
Manual Google → copy/paste → write yourself	AI clarifies your needs → deep multi-perspective search → auto-generation
Single viewpoint	🆕 3–5 perspectives (policy/market/tech/social/international)
Text-only reports, no visuals	🆕 ECharts charts + Mermaid diagrams
One-shot generation, can't update	🆕 Dynamic supplement loop: paste links → integrate → regenerate
Ugly Word formatting	Typst academic-grade PDF: cover + TOC + pagination + citations

Features

• 🎯 Smart Requirement Clarification — 4-dimension structured interview. ~10–15 rounds, with a progress bar so users know what to expect.
• 🔍 STORM Multi-Perspective Search — Inspired by the Stanford paper. Discovers 3–5 viewpoints, each explored in depth.
• 💬 Simulated Expert Dialogue — 3 rounds per perspective: initial search → follow-up → deep dive.
• ✅ Fact Verification — AI extracts key claims, cross-references ≥2 independent sources, labels "verified / single-source / disputed".
• 📊 Auto Chart Generation — ECharts for data + Mermaid for flowcharts/timelines.
• 📄 Academic-Grade PDF — Typst engine: cover, TOC, pagination, citations, methodology section, limitations disclosure.
• 🔄 Dynamic Supplement Loop — After generation, paste new links → auto-fetch → merge → regenerate. Repeat as needed.
• 💾 Session Persistence — Save, list, delete, resume research sessions anytime.

Quick Start

Prerequisites: Node.js 16+, Python 3.8+, Typst v0.14+, DeepSeek API Key

git clone https://github.com/Rookage/deep-research-assistant.git
cd deep-research-assistant
npm install
pip install httpx lxml python-docx python-pptx

# Install Typst
choco install typst        # Windows
brew install typst          # macOS

cp .env.example .env        # Add your DeepSeek key
npm start                   # http://localhost:3001

Architecture

Vague Idea
    │
    ▼
┌──────────────────────────┐
│ Stage 1: Clarify + Views  │  ← DeepSeek + STORM
│ Interview → Brief → 3–5   │
│ research perspectives     │
└────────┬─────────────────┘
         │ ✅ User confirms
         ▼
┌──────────────────────────┐
│ Stage 2: Deep Search      │  ← Baidu + Bing + httpx/lxml
│ Per-perspective 3-round   │
│ expert dialogue           │
└────────┬─────────────────┘
         │ ✅ User confirms
         ▼
┌──────────────────────────┐
│ Stage 3: Fact Check       │
│ AI extracts claims →      │
│ cross-verify → label      │
└────────┬─────────────────┘
         │ ✅ User confirms
         ▼
┌──────────────────────────┐
│ Stage 4: Report Generation│  ← Typst PDF / DOCX / PPTX
│ Outline → sections+charts │
│ → polish → download       │
│ 🔄 Supplement Loop        │
└──────────────────────────┘

Tech Stack

Layer	Tech
Backend	Node.js + Express
AI	DeepSeek API
Search	Baidu + Bing
PDF	Typst v0.14.2
Office	python-docx + python-pptx
Charts	ECharts + Mermaid
Content Fetch	httpx + lxml (Python)
Frontend	Vanilla HTML/CSS/JS
Methodology	Stanford STORM

License

MIT

---

🇨🇳 中文

为什么你需要这个工具？

传统做法	本工具
手动 Google → 复制粘贴 → 自己写	AI 自动追问澄清需求 → 多视角深度搜索 → 自动生成
只看一个角度的信息	🆕 3-5 个视角（政策/市场/技术/社会/国际）
报告只有字，没有图	🆕 ECharts 图表 + Mermaid 流程图
生成完就不能改了	🆕 动态补充 Loop：贴链接 → 整合 → 再生
Word 排版惨不忍睹	Typst 学术级 PDF：封面+目录+页码+引用

核心特性

• 🎯 智能需求澄清 — 4 维度结构化访谈，预计 10~15 轮，前端有进度条提示
• 🔍 STORM 多视角搜索 — 借鑑 Stanford 论文，自动发现 3-5 个研究视角
• 💬 模拟专家对话 — 每个视角 3 轮追问：初始搜索 → 追問 → 深入追問
• ✅ 事实交叉验证 — AI 识别关键主张，至少 2 个独立来源确认
• 📊 图表自动生成 — ECharts 数据图表 + Mermaid 流程图/时间线
• 📄 学术级 PDF — Typst 引擎，封面+目录+页码+引用+方法论+局限性声明
• 🔄 动态补充 Loop — 报告生成后贴入链接 → 自动抓取 → 整合进报告 → 重新生成
• 💾 会话持久化 — 支持恢复、列表、删除，随时继续之前的研究

快速开始

前置要求： Node.js 16+, Python 3.8+, Typst v0.14+, DeepSeek API Key

git clone https://github.com/Rookage/deep-research-assistant.git
cd deep-research-assistant
npm install
pip install httpx lxml python-docx python-pptx

# 安装 Typst
choco install typst        # Windows
brew install typst          # macOS

cp .env.example .env        # 填入 DeepSeek API Key
npm start                   # http://localhost:3001

四阶段流水线

输入模糊需求 → 视角发现 → 多视角深度搜索 → 事实核查 → 学术级报告
                                                                  ↕
                                                           🔄 补充链接 Loop

技术栈

层	技术
后端	Node.js + Express
AI	DeepSeek API
搜索	百度 + Bing
PDF	Typst v0.14.2
图表	ECharts + Mermaid
内容抓取	httpx + lxml（Python）
方法论	Stanford STORM

许可证

MIT

---

_{Built with ❤️ by Rookage}