pdf-reader

name	PDF Reader
description	"深度阅读和分析 PDF 论文。支持 PDF 转 Markdown、智能摘要、关键信息提取、问答式学习、笔记生成。"

When to Use

Use this skill when you want to:

• Read and analyze PDF papers in depth
• Extract key information from PDFs
• Convert PDF to Markdown
• Ask questions about a paper
• Generate comprehensive notes

Triggers:

• User says "read this PDF", "read paper", "阅读这篇论文"
• User provides a PDF file or path
• User asks "analyze this paper", "分析这篇论文"
• User mentions "PDF to Markdown", "PDF 转 Markdown"
• User wants "paper summary", "论文摘要"

Quick Start

Read a PDF

阅读 PDF: D:\Papers\attention-is-all-you-need.pdf

Read from arXiv

阅读 arXiv 论文: 2303.12345

Ask Questions

这篇论文的主要贡献是什么？

Manual Run

python {workspace}/PaperVault/scripts/pdf-reader.py --pdf path/to/paper.pdf

Features

1. PDF to Markdown Conversion

Converts PDF to structured Markdown:

Input: PDF file Output: {paper-name}.md

Conversion includes:

• Text extraction
• Section detection
• Figure/table references
• Equation formatting
• Reference linking

Example Output:

---
title: "Attention Is All You Need"
arxiv_id: "1706.03762"
date_read: 2024-03-25
reading_time: 45min
difficulty: ⭐⭐⭐☆☆
---

# Attention Is All You Need

> [!abstract] 摘要
> The dominant sequence transduction models are based on complex recurrent or convolutional neural networks...

## 1. Introduction

Recurrent neural networks, long short-term memory and gated recurrent neural networks...

### 1.1 Background

The Transformer uses multi-headed self-attention...

## 2. Model Architecture

![Model Architecture](figures/model-architecture.png)

### 2.1 Encoder and Decoder Stacks

**Encoder**: The encoder is composed of a stack of N = 6 identical layers...

**Decoder**: The decoder is also composed of a stack of N = 6 identical layers...

## 3. Attention

$$
\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V
$$

## 4. Experiments

| Model | BLEU | Training Time |
|-------|------|---------------|
| Transformer (big) | 28.4 | 3.5 days |
| Transformer (base) | 27.3 | 12 hours |

## Key Insights

1. Self-attention allows modeling of dependencies regardless of distance
2. Multi-head attention enables attending to information from different positions
3. Positional encoding is necessary since the model contains no recurrence

## Questions

- [ ] Why divide by sqrt(d_k)?
- [ ] How does multi-head attention work in detail?
- [ ] What are the computational complexity trade-offs?

## References

1. Vaswani et al. (2017) - This paper
2. Bahdanau et al. (2015) - Attention mechanism
3. Gehring et al. (2017) - Convolutional sequence models

2. Intelligent Summarization

Generates multiple levels of summaries:

1-sentence summary: "Proposes Transformer, a purely attention-based architecture for sequence transduction."

1-paragraph summary: "The paper introduces the Transformer model, which relies entirely on self-attention mechanisms without recurrence or convolution. It achieves state-of-the-art performance on machine translation tasks with significantly reduced training time."

Key points summary:

• Novel architecture based solely on attention
• Eliminates recurrence and convolution
• Achieves new SOTA on translation tasks
• More parallelizable and faster to train

3. Information Extraction

Extracts structured information:

Contributions:

{
  "contributions": [
    "Proposed Transformer architecture",
    "Introduced multi-head self-attention",
    "Achieved SOTA on WMT 2014",
    "Reduced training time significantly"
  ]
}

Methods:

{
  "methods": [
    {
      "name": "Scaled Dot-Product Attention",
      "equation": "Attention(Q,K,V) = softmax(QK^T / sqrt(d_k))V",
      "purpose": "Compute attention weights efficiently"
    },
    {
      "name": "Multi-Head Attention",
      "description": "Run multiple attention operations in parallel",
      "benefit": "Attend to information from different positions"
    }
  ]
}

Experiments:

{
  "experiments": [
    {
      "task": "Machine Translation",
      "dataset": "WMT 2014 English-German",
      "metric": "BLEU",
      "result": 28.4,
      "baseline": 26.4
    }
  ]
}

4. Q&A Mode

Interactive question answering:

User: "What is the main innovation?" Assistant: "The main innovation is replacing recurrent layers with self-attention mechanisms, allowing for more parallelization and better modeling of long-range dependencies."

User: "How does self-attention work?" Assistant: "Self-attention computes relationships between all positions in a sequence simultaneously. For each position, it computes attention weights with respect to all other positions..."

User: "What are the limitations?" Assistant: "Based on the paper and subsequent research: 1) Quadratic memory complexity for long sequences, 2) Position encoding may not generalize well, 3) Limited to fixed-length contexts..."

5. Note Generation

Generates comprehensive notes:

Concept Notes:

# Self-Attention

## Definition
Self-attention is a mechanism that relates different positions of a single sequence...

## How It Works
1. Compute Query, Key, Value matrices
2. Calculate attention weights
3. Apply weights to values
4. Output weighted sum

## Why It Matters
- Captures long-range dependencies
- Highly parallelizable
- Interpretable attention patterns

## Applications
- Machine translation
- Text summarization
- Image generation

## Related Concepts
- [[Attention Mechanism]]
- [[Transformer]]
- [[Multi-Head Attention]]

Method Notes:

# Multi-Head Attention

## Overview
Runs multiple self-attention operations in parallel...

## Implementation
```python
class MultiHeadAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super().__init__()
        self.num_heads = num_heads
        self.d_k = d_model // num_heads
        # ...

Advantages

• Different heads can focus on different aspects
• More expressive than single attention
• Still computationally efficient

Trade-offs

• More parameters
• Requires careful tuning
• May overfit on small datasets

## Reading Workflow

### Level 1: Quick Scan (5-10 min)
- Title and abstract
- Introduction
- Conclusion
- Key figures

**Output**: High-level understanding

### Level 2: Standard Read (30-60 min)
- All sections
- Important equations
- Key experiments
- Method details

**Output**: Detailed notes + questions

### Level 3: Deep Dive (2-4 hours)
- Every section in detail
- Derive equations
- Reproduce experiments
- Related work

**Output**: Comprehensive understanding + implementation

## Advanced Features

### Batch Processing
Process multiple PDFs:
```bash
python pdf-reader.py --dir D:\Papers\ --batch

Comparison Mode

Compare multiple papers:

对比分析这两篇论文的异同

Export Options

• Markdown: Default format
• HTML: With styling
• PDF: Annotated version
• JSON: Structured data

Integration with Zotero

Read from Zotero

python pdf-reader.py --zotero --query "attention mechanism"

Save Back to Zotero

python pdf-reader.py --pdf paper.pdf --save-to-zotero

Customization

Custom Prompts

Create prompts/custom-summary.txt:

请从以下角度总结这篇论文：
1. 研究动机
2. 核心创新
3. 技术方案
4. 实验验证
5. 局限性

Reading Templates

Create templates/custom-template.md:

# {title}

## 一句话总结
{one_sentence_summary}

## 核心观点
{key_points}

## 技术细节
{technical_details}

## 我的思考
<!-- Your reflections -->

Troubleshooting

PDF Extraction Failed

Error: Cannot extract text from PDF

Solutions:

1. PDF might be scanned image - needs OCR
2. Try different extraction tool
3. Use --ocr flag

Poor Quality Conversion

Warning: Many formatting errors

Solutions:

1. Check PDF quality
2. Use --enhance flag
3. Manual review needed

LLM Rate Limit

Error: API rate limit exceeded

Solution: Wait or use different API key

Best Practices

1. Start with Abstract: Always read abstract first
2. Active Reading: Take notes while reading
3. Ask Questions: Don't skip confusing parts
4. Connect Knowledge: Link to related papers
5. Review Later: Re-read important papers

Keyboard Shortcuts (when using interactive mode)

Key	Action
n	Next section
p	Previous section
q	Ask question
s	Save note
h	Highlight text
f	Find in paper
Esc	Exit

Related Skills

• paper-fetcher - Get papers to read
• paper-summarizer - Generate summaries
• learning-reflector - Reflect on reading
• zotero-local - Manage PDF library

Future Enhancements

• Audio reading (TTS)
• Collaborative annotation
• Citation network visualization
• Automated fact-checking
• Multi-language support