Skip to content

GoDiao/Paper-Reader

BranchesTags

Repository files navigation

Paper Reader Agent

Paper Reader Agent Banner

Python 3.8+ License: AGPL-3.0 DeepSeek OpenAI Tavily Valyu

Hierarchical Multi-Agent System for Academic Paper Deep Analysis & AI-Powered Web Research

中文文档 | English

Quick Links: Paper Reader | Deep Research | Documentation


📖 Introduction

Paper Reader Agent is an advanced AI system designed to read, analyze, and synthesize academic papers with a depth that matches human researchers.

Deep Research is a powerful AI-powered web research tool that leverages Tavily and Valyu APIs to conduct comprehensive, real-time research on any topic, generating detailed reports with citations from online sources.

Together, they form a complete research workflow: Paper Reader for deep paper analysis, and Deep Research for broad topic exploration.

📸 Click to see Screenshots (UI & Features)

Paper Reader

Modern Web UI (Bilingual) Real-time Progress Tracking
English UI Analysis Progress
Clean interface with EN/ZH switching Visualize the 5-agent team in action
Publication-Quality Reports Specialist Deep Dives
Final Report Specialist Reports
Auto-embedded figures & formulas Rich details from specific domains

Deep Research

Independent Research Page Real-time Streaming Results
Deep Research UI Research Processing
Clean, focused research interface Live streaming with progress tracking
Research Dashboard
Research Dashboard
Manage research history, export to Notion

Paper Reader Agent goes beyond simple summarization...

Unlike standard summary tools, it employs a Hierarchical Multi-Agent Architecture (1+3+1) to mimic a professional research team:

  1. Architect: Deconstructs the paper and plans the reading strategy.
  2. Specialist Team: Parallel experts analyze Context, Math, and Data.
  3. Editor: Synthesizes a publication-quality report with embedded figures.

Key Feature: The system detects, extracts, and literally sees figures, embedding them directly into the analysis where they are discussed, maintaining full visual context.

🏗️ Architecture

The system operates using a "Divide and Conquer" strategy orchestrated by a central planner.

Architecture Diagram

✨ Key Features

🧠 Paper Reader - Hierarchical Intelligence

  • Architect Agent: Strategic planning of what to read and where to focus.
  • Context Hunter: Digs for the "real" motivation and hidden assumptions.
  • Math Specialist: Derives equations and explains physical intuition behind formulas.
  • Data Auditor: Critically checks baselines, variance, and experimental fairness.

👁️ Paper Reader - Visual Understanding

  • Smart Extraction: Custom PDF parsing pipeline (based on PyMuPDF) that segments text and images.
  • Context Preservation: Figures are kept with their relevant text.
  • Auto-Embedding: The AI inserts figures into the report exactly when discussing them.

🌐 Deep Research - Web-Powered Research

  • Dual Provider Support: Choose between Tavily (fast, reliable) and Valyu (comprehensive, multi-tier).
  • Real-time Streaming: Watch research progress live with SSE streaming technology.
  • Citation Management: Multiple citation formats (Numbered, APA, MLA, Chicago).
  • Research History: Persistent storage with browse, search, and delete capabilities.
  • Export to Notion: One-click export with full formatting, tables, and LaTeX equations.

💻 Modern Interaction

  • Web Interface: Clean, responsive UI with real-time analysis progress.
  • Dual-Mode:
    • Simple: Quick architect + math check.
    • Hierarchical: Full 5-agent deep dive.
  • Bilingual: Generates native-quality English and Chinese reports simultaneously.
  • Independent Research Page: Dedicated /researcher page for Deep Research with isolated history.

🚀 Quick Start

Prerequisites

  • Python 3.8+
  • API Key (DeepSeek or OpenAI)
  • (Optional) CUDA GPU for faster layout analysis

Installation

git clone https://github.com/GoDiao/Paper-Reader.git
cd Paper-Reader
python -m venv .venv

# macOS / Linux
source .venv/bin/activate
# Windows (PowerShell)
.venv\Scripts\Activate.ps1

pip install -r requirements.txt

Optional dependencies:

# PDF export (Windows needs extra system dependencies; see weasyprint docs)
pip install weasyprint

Configuration

Copy the env template and fill in your keys (do not commit .env):

# macOS / Linux
cp .env.example .env

# Windows (PowerShell)
Copy-Item .env.example .env

Minimal .env (Paper Reader):

DEEPSEEK_API_KEY=sk-your-key
# OR
OPENAI_API_KEY=sk-your-key

Deep Research Configuration (add to .env):

# Tavily API (required for Deep Research)
TAVILY_API_KEY=tvly-your_api_key_here

# Valyu API (alternative provider for Deep Research)
VALYU_API_KEY=your_valyu_api_key_here

# Notion Export (optional)
NOTION_SECRET=your_notion_secret
NOTION_PARENT_PAGE_ID=your_parent_page_id
IMGBB_API_KEY=your_imgbb_api_key

Optional settings (web search enrichment + extra providers):

SILICONFLOW_API_KEY=your_siliconflow_api_key_here
ENABLE_WEB_SEARCH=false
GITHUB_TOKEN=your_github_token_here
HUGGINGFACE_TOKEN=your_huggingface_token_here
SERPER_API_KEY=your_serper_api_key_here

Usage

1. Web Interface (Recommended)

Start the server to enjoy the full interactive experience.

python web_server.py

Open http://localhost:8000 in your browser for Paper Reader, or visit http://localhost:8000/researcher for Deep Research.

Note (Simple mode in Web UI)
The Simple mode in the web UI is currently a placeholder and returns a brief message. Use Hierarchical mode for full reports.

Note (Parser Backend in Web Mode)
The web server currently uses the auto strategy by default:

  • If MinerU (pip install mineru) is installed, it will try the MinerU backend first.
  • If MinerU is not installed or fails, it will automatically fall back to the PyMuPDF backend.

2. Deep Research - Web Interface

Access the dedicated Deep Research page at http://localhost:8000/researcher:

  1. Enter your research topic or question
  2. Select provider (Tavily or Valyu)
  3. Choose model and citation format
  4. Click "Start Research" and watch real-time streaming
  5. Save, export to Notion, or manage history

Deep Research Features:

  • ✅ Real-time SSE streaming with progress tracking
  • ✅ Markdown rendering with tables and LaTeX equations
  • ✅ Persistent research history with search and delete
  • ✅ One-click export to Notion with full formatting

3. Command Line

# Full hierarchical analysis (Default, auto parser backend)
python main.py paper.pdf

# Force fast PyMuPDF backend
python main.py paper.pdf --parser pymupdf

# Force high-fidelity MinerU backend (requires: pip install mineru)
python main.py paper.pdf --parser mineru

# Save intermediate agent outputs
python main.py paper.pdf --verbose

# Use OpenAI instead of DeepSeek
python main.py paper.pdf --provider openai --model gpt-4o

# Output Chinese only (can reduce cost)
python main.py paper.pdf --language zh

Note: Deep Research is currently only available through the web interface (/researcher). CLI support is planned for future releases.

PDF Parsers: PyMuPDF vs MinerU

  • PyMuPDF (Default, Fast)

    • No extra dependencies beyond pymupdf.
    • Very fast, good enough for most standard papers.
    • Enhanced in this project with table extraction, math-region heuristics, and smarter figure detection.

  • MinerU (Optional, High-Fidelity)

    • Install via pip install mineru (and follow MinerU's own docs for GPU/driver requirements).
    • Better at preserving complex layouts, multi-column structure, tables, and math-heavy pages.
    • When used, this project normalizes MinerU's Markdown + images into the same ParsedDocument format as PyMuPDF, so downstream agents and UI work identically.

Repository Note
This project supports pip install mineru as an optional parsing backend; MinerU manages its own model cache (usually under your user/cache directory).
This repository also contains the upstream MinerU/ source tree (licensed under AGPL-3.0). If you want a permissive license for your app code, avoid shipping MinerU source in the same repo.

📂 Output Structure

Paper Reader (Web mode: python web_server.py)

outputs/
└── {upload_id}/
    ├── paper_analysis.md
    ├── paper_analysis_zh.md
    ├── images/
    ├── specialists/
    └── figure_index.json

data/
└── reports.json

Paper Reader (CLI mode: python main.py ...)

output/
└── {pdf_stem}/
    ├── paper_analysis.md
    ├── paper_analysis_zh.md
    ├── images/
    ├── parsed/
    ├── specialists/
    └── figure_index.json

Deep Research (Web mode: /researcher)

data/
└── researches.json  # Independent research history storage

Research results are stored separately from paper analysis reports and include:

  • Full Markdown content with tables and LaTeX equations
  • Source list with titles, URLs, and favicons
  • Model and citation format metadata
  • Creation and update timestamps

🛠️ Project Structure

paper_reader/
├── agents/                 # 🤖 The Brains
│   ├── hierarchical_orchestrator.py
│   ├── hierarchical_prompts.py
│   └── ...
├── parsers/                # 👁️ The Eyes
│   └── pdf_parser.py       # Custom Layout Analysis
├── generators/             # 📝 The Scribe
│   └── report_generator.py # Report Assembly
├── backend/                # 🔌 API Server
│   ├── app.py              # Main FastAPI application
│   ├── research_store.py   # Deep Research storage
│   ├── deep_research_utils.py  # Deep Research utilities
│   └── ...
├── frontend/               # 🖥️ Web UI
│   ├── index.html          # Paper Reader UI
│   └── researcher.html     # Deep Research UI (Independent page)
├── services/               # 🌐 External Services
│   ├── tavily_service.py   # Tavily Deep Research API wrapper
│   ├── valyu_service.py    # Valyu Deep Research API wrapper
│   └── ...
└── deep_research/          # 📚 Documentation
    ├── tavily/             # Tavily API documentation
    └── valyu/              # Valyu API documentation

🚀 Changelog

v2.0.0 - Deep Research Integration

🎉 Major Addition: Deep Research - AI-Powered Web Research Tool

  • 🌐 Deep Research Feature:

    • Independent research page at /researcher with dedicated UI
    • Dual provider support: Tavily (fast, reliable) and Valyu (comprehensive, multi-tier)
    • Real-time SSE streaming with live progress tracking
    • Multiple citation formats: Numbered, APA, MLA, Chicago
    • Persistent research history with search and delete
    • One-click export to Notion with full formatting

  • 🔧 Backend Infrastructure:

    • ResearchStore for independent research storage
    • TavilyService and ValyuService wrappers
    • Unified polling and progress tracking
    • Structured error handling with code/details

  • 🖥️ Frontend Features:

    • Dedicated researcher.html page with modern glassmorphism design
    • Real-time Markdown rendering with tables and LaTeX equations
    • Research history panel with isolated storage
    • Export to Notion with native tables and equations

  • 📊 API Endpoints:

    • POST /api/deep-research - Start research via WebSocket
    • GET /api/deep-research/stream - SSE streaming endpoint
    • GET /api/research - List research history
    • POST /api/research/save - Save research
    • DELETE /api/research/{id} - Delete research
    • POST /api/research/{id}/export/notion - Export to Notion

  • 📚 Documentation:

    • Comprehensive API documentation for Tavily and Valyu
    • Streaming implementation guide
    • Independent page design documentation

v1.7.0 - Iterative Analysis & Gap Agent

🎉 Major Announcement: Starting from v1.7.0, Paper Reader officially supports one-click export to Notion!

  • 🔄 Iterative Analysis:

    • Specialists can proactively raise information gaps (<TENTATIVE_GAPS>), triggering automatic refinement cycles.
    • Configurable iteration rounds (max_iterations), with each round resolving requests from the previous iteration.
    • New frontend iteration panel displaying request type, content, and resolution status per round.

  • 🧩 Gap Agent (Gap Analysis Specialist):

    • New independent Gap Agent that reviews all three specialist reports to identify cross-domain information gaps.
    • Generates global confidence score (0.0–1.0) with natural language explanation for iteration recommendations.
    • Outputs standardized unified_requests list, auto-classified as section_needed, cross_reference, clarification, or figure_detail.

  • 📊 Frontend Enhancements:

    • Gap Agent Panel: Displays per-specialist assessments (completeness, coherence, gaps found), global confidence, iteration recommendation, and request list.
    • Round Tracking Fix: WebSocket events now carry correct round numbers, displaying "Round N" accurately.
    • Config Display Fix: max_iterations=0 no longer incorrectly shows as "2".

  • 🛠️ Backend Optimizations:

    • Round Logic Fix: max_iterations=N now truly executes N+1 specialist rounds (initial + N refinements).
    • JSON Parsing Enhancement: Added json-repair fallback to handle unescaped quotes, newlines, and other malformed JSON from LLMs.
    • Cache Optimization: Empty parse results (0 characters) are no longer cached to prevent pollution.
    • MinerU Compatibility: Handles Magika returning unknown for PDF identification, improving success rate for complex PDFs.

  • 📦 New Dependency:

    • json-repair>=0.55.0: Automatically repairs malformed JSON from LLM outputs.

tonotion (tag) - Notion Export (Side Feature)

  • 📝 Native Notion Export: One-click export to Notion pages with full support for:
    • Native Tables: Clean, editable Notion tables (replacing image-based or LaTeX tables).
    • Rich Math Support: Perfect rendering of inline ($...$) and block ($$...$$) equations.
    • Nested Lists: Correct indentation for complex nested lists.
  • 🖼️ Image Optimization: Improved figure resolution and caption handling during export.
  • ⚡ Core Stability: Fixed edge cases in list parsing and table generation.

v1.6.0 - History, Export & Performance

  • 🗂️ Report History: Persistent report store with browse/search/delete, plus reloadable specialist reports and chat history.
  • 📤 One-click Export: Export reports as Markdown/DOCX, download extracted figures as a ZIP (PDF export supported via optional dependencies).
  • ⚡ PDF Parse Cache: SHA256-based parse caching (parsed content + figures) to significantly speed up repeated analyses.
  • 🌐 Web Search Toggle: UI switch to enable/disable reproduction resource discovery, with GitHub/HuggingFace token support.
  • 🤝 Provider Expansion: Added SiliconFlow provider support in the web UI with concurrency tuning to reduce rate-limit errors.
  • 🈯 Output Language Control: Choose EN or ZH output to reduce cost and avoid empty report tabs.

v1.5.0 - UI Modernization & Resource Discovery

  • 🎨 Modern UI Overhaul: Complete redesign with a new Zinc-based dark theme, high-contrast tables for better readability, and refined typography.
  • 🔍 Resource Discovery Services: Integrated automated search for reproduction resources (GitHub code repositories, HuggingFace models/datasets) directly into the analysis pipeline.
  • 📋 Reproduction Checklist: New dedicated section to extract and verify hardware requirements, hyperparameters, and datasets.
  • 📉 Variable Tracking: Added support for tracking mathematical variables and their definitions across the paper.
  • ⚡ UX Refinements: Streamlined the agent progress view by removing the redundant Architect tab, focusing on the specialist analysis.

v1.3.0 - MinerU Parsing Upgrade

  • 🧠 MinerU Parser Backend: Integrated MinerU (Magic-PDF 2.x pipeline) as a high-fidelity PDF parser for complex academic papers, with better layout, table, and math structure preservation.
  • ⚙️ Switchable PDF Backend: Added a selectable parser backend in CLI (--parser auto|pymupdf|mineru) and web mode, so you can choose fast PyMuPDF, high-quality MinerU, or an auto strategy that tries MinerU first and falls back to PyMuPDF if unavailable or failing.
  • 📂 Unified Output Pipeline: Normalized MinerU outputs into the existing ParsedDocument + figure index flow so that downstream LLM agents, report generation, and UI work seamlessly regardless of which parser backend you choose.

v1.2.0 - Architecture & Performance Improvements

  • 🔧 Unified LLM Client Factory: Centralized LLM configuration, retry/backoff, and timeout handling across all agents. Added optional global concurrency limiting to prevent API rate limits.
  • 📊 Fine-grained Progress Events: Real-time progress updates for each agent (Architect, Context Hunter, Math Specialist, Data Auditor, Editors) with detailed status messages during LLM calls and retries.
  • ⚡ Concurrency Optimization: Eliminated nested thread pools, unified executor management, and improved resource utilization for better performance under concurrent loads.
  • 📄 Enhanced PDF Parser: Improved PyMuPDF implementation with table extraction (Markdown format), mathematical formula region detection, better text structure preservation, and smarter image caption detection (searches above/below images).
  • 🔧 Configuration: New environment variables (LLM_TIMEOUT_S, LLM_MAX_RETRIES, LLM_MAX_CONCURRENCY) for fine-tuning API behavior.
  • 📡 Real-time Streaming: Implemented streaming responses for expert agents (Math, Data, Context), allowing users to see reports generating token-by-token.
  • 📐 Math Formula Fix: Solved critical rendering issues for streamed LaTeX formulas by protecting delimiters (\[...\], \(...\)) from Markdown processing.
  • 🏗️ Architect Report: Added a new dedicated "Architect" tab to visualize the reading plan and agent assignments immediately after the planning phase.
  • 🖥️ UI UX Improvements: Moved Specialist Reports to the main view for better visibility and added auto-focus logic to follow the active agent.

v1.1.0 - Premium UI & Chat Upgrade

  • ✨ New UI: Introduced a "Deep Space" glassmorphism theme for a premium reading experience.
  • 🤖 Smart Chat: Added context summarization to the Chat AI, allowing for longer, more coherent discussions about the paper.
  • 📊 Specialist Reports: View detailed analysis from specific agents (Context Hunter, Math Specialist, Data Auditor) in dedicated tabs.
  • 🌐 Bilingual: Added full English/Chinese language switching.
  • 🐛 Fixes: Resolved table rendering issues and improved chat interface scrolling.

📄 Documentation

For detailed information about Deep Research features:

🤝 Contributing

Contributions are welcome! Whether it's a new specialist agent, better parsing logic, or UI improvements.

  1. Fork the Project
  2. Create your Feature Branch
  3. Commit your Changes
  4. Push to the Branch
  5. Open a Pull Request

📄 License

This repository includes MinerU/ (AGPL-3.0), so redistribution must follow AGPL-3.0. See LICENSE.md.

Star ⭐ this repo if it helped your research!

About

An AI-powered academic paper analysis system with hierarchical multi-agent architecture and interactive web interface.

Resources

Readme
Activity

Stars

4 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

8 tags

Packages

 
 
 

Contributors

Languages