Academic Review

Document Type Detection

When starting a review session, identify the document type from context:

Lectures - Indicators:

• File names with "lecture", "slides", "presentation"
• Bullet-point heavy content
• Sequential slide numbers
• Course/semester codes (e.g., "CS229_Lecture05.pdf")

Research Papers - Indicators:

• File names with "paper", author names, conference/journal codes
• Standard sections: Abstract, Introduction, Methods, Results, Discussion, References
• Citations and bibliography
• Two-column format common

Book Chapters - Indicators:

• File names with "chapter", book titles
• Sections and subsections with numbered headings
• End-of-chapter exercises or problems
• Dense paragraph-based text

Default approach: If unclear, start with Q&A mode and adapt based on the content structure.

Quick Start

Starting a Review Session

When the user requests a review session:

1. Find PDFs using Glob: **/*.pdf or more specific patterns
2. Identify document type (lecture/paper/chapter) from filename and request
3. Extract content using the extraction script
4. Ask mode preference: "Would you like Q&A mode or Quiz mode?"
5. Begin the selected mode with document-type-appropriate approach

Example Flow

User: "Review the SLAM paper by Smith et al."

Your actions:
1. Use Glob to find PDFs matching "smith" or "slam"
2. Identify as research paper
3. Run: python scripts/extract_pdf.py <pdf_path>
4. Read the cached markdown file
5. Ask: "Q&A mode or Quiz mode?"
6. Begin selected mode (adapt to paper structure)

Review Modes

Q&A Mode (Free-Form Questions)

Purpose: Answer user's specific questions about content with detailed explanations.

How to conduct Q&A mode:

1. Load and parse content:

   # Extract PDF (or use cached version)
   python scripts/extract_pdf.py /path/to/document.pdf
   

   Then read the output markdown file using the Read tool.

2. Present overview (adapt to document type):

   For Lectures:

   • List main topics covered
   • Highlight key formulas (show in LaTeX)
   • Mention important definitions or concepts

   Example:

   📚 Lecture Overview: Epipolar Geometry
   
   This lecture covers 45 slides on:
   - Epipolar constraint: $x'^T F x = 0$
   - Fundamental matrix $F$ (3x3, rank 2)
   - Essential matrix $E = K'^T F K$
   - Applications: stereo vision, 3D reconstruction
   
   Ask me anything about these topics, or say "quiz" to switch to quiz mode.
   

   For Research Papers:

   • Summarize the research question/contribution
   • Key methodology and approach
   • Main results and conclusions
   • Important formulas or algorithms

   Example:

   📄 Paper Overview: "ORB-SLAM2: Real-Time SLAM for Monocular, Stereo and RGB-D Cameras"
   
   **Research Question**: How to build a complete SLAM system that works across multiple camera types?
   
   **Key Contributions**:
   - Unified SLAM system for monocular, stereo, and RGB-D cameras
   - Place recognition and loop closing
   - Real-time performance on standard CPUs
   
   **Methods**: ORB features, bag-of-words place recognition, pose graph optimization
   
   **Results**: Evaluated on KITTI and TUM datasets, outperforms previous methods
   
   Ask me about methodology, results, or implementation details.
   

   For Book Chapters:

   • Main concepts introduced
   • Theorems or key results
   • Important formulas
   • Example problems covered

   Example:

   📖 Chapter Overview: "Matrix Decompositions" (Chapter 7)
   
   **Topics Covered**:
   - Singular Value Decomposition (SVD): $A = U\Sigma V^T$
   - Eigenvalue decomposition: $A = Q\Lambda Q^T$
   - QR decomposition and applications
   - Least squares via matrix decompositions
   
   **Key Theorems**: Spectral theorem, SVD existence
   
   **Exercises**: 15 problems on computing decompositions and applications
   
   Ask me about concepts, work through examples, or get help with exercises.
   

Document-specific guidance:

Lectures: Focus on concept understanding, derivations, applications Papers: Focus on methodology critique, results interpretation, reproducibility Chapters: Focus on theorem understanding, example walkthrough, exercise solving

Quiz Mode (Auto-Generated Questions)

Purpose: Test user's knowledge with auto-generated questions, provide scoring and feedback.

How to conduct Quiz mode:

1. Load and analyze content:

   # Extract PDF (or use cached version)
   python scripts/extract_pdf.py /path/to/document.pdf
   

   Read the markdown and analyze:

   • Key concepts and definitions
   • Important formulas (in LaTeX)
   • Learning objectives
   • Example problems

2. Generate questions (default: 10, but ask user for preference):

   For Lectures:

   • Multiple choice: Test understanding of concepts
   • True/False: Quick concept checks
   • Short answer: Define terms or explain relationships
   • Formula problems: Apply equations to scenarios

   For Research Papers:

   • Multiple choice: Methodology choices, experimental design
   • True/False: Claims about results or methods
   • Short answer: Explain key contributions, limitations
   • Analysis questions: Critique methods or interpret results

   For Book Chapters:

   • Multiple choice: Theorem conditions, concept understanding
   • True/False: Mathematical statements
   • Short answer: Prove simple results, explain concepts
   • Problems: Similar to end-of-chapter exercises

   Mix question types and topics proportionally. Order by difficulty (easier first).

3. Present questions one at a time:

   Quiz Mode - 10 Questions
   Score: 0/0
   
   Question 1 of 10 [Multiple Choice]
   What is the rank of the Fundamental matrix $F$?
   
   a) 1
   b) 2
   c) 3
   d) 4
   
   Your answer:
   

4. Evaluate and provide feedback:

   ✓ Correct! [+1 point]
   
   The Fundamental matrix $F$ has rank 2, which means det($F$) = 0. This constraint
   arises from the fact that $F$ maps points to epipolar lines, and the mapping has
   a one-dimensional null space.
   
   Score: 1/1 (100%)
   
   Question 2 of 10...
   

Scoring Guidelines:

• Multiple choice: 1 point for correct answer
• True/False: 1 point for correct answer
• Short answer: 1 point if answer captures key concept (be flexible)
• Formula problems: 1 point for correct answer, 0.5 for correct approach but calculation error

Finding PDFs

General patterns:

# All PDFs in current directory and subdirectories
glob pattern: "**/*.pdf"

# Find specific document by name
glob pattern: "**/*smith*.pdf"

# Course-specific (if organized in directories)
glob pattern: "CS229/**/*.pdf"

When user's request is ambiguous:

1. Use Glob to find matching PDFs
2. Present options if multiple matches
3. Let user select which PDF to review

Extraction and Caching

First-Time Extraction

When extracting a PDF for the first time:

# Run extraction script (uses venv python)
source .venv/bin/activate && python scripts/extract_pdf.py /path/to/document.pdf

This will:

• Convert PDF to markdown using Marker
• Preserve formulas as LaTeX
• Cache result in .cache/extracted/<hash>.md
• Save metadata in .cache/extracted/<hash>.json
• Take 30-60 seconds depending on PDF size

The script prints the path to the cached markdown file - use Read to load it.

Using Cached Extraction

If the PDF was previously extracted:

• Script immediately returns cached file path
• Read the markdown file using Read tool
• Instant access (no re-processing)

Extracting Page Images (Fallback)

If user asks about a specific diagram or visual element:

# Extract page 23 as an image
source .venv/bin/activate && python scripts/extract_page_image.py /path/to/document.pdf 23

Then use Read to view the image and analyze it visually.

Working with LaTeX Formulas

Extracted formulas are in LaTeX format:

Inline: $E = mc^2$ Display: $$\int_a^b f(x)dx$$

When explaining formulas:

• Show the LaTeX notation
• Explain each variable
• Provide context from the document
• Give examples if helpful

Example:

The epipolar constraint is expressed as:

$$x'^T F x = 0$$

Where:
- $x$ and $x'$ are corresponding points in homogeneous coordinates
- $F$ is the 3×3 Fundamental matrix
- The equation states that $x'$ lies on the epipolar line $Fx$ in the second image

Tips for Best Results

General Guidelines:

1. Always extract first: Run the extraction script before answering questions
2. Use caching: Check if extraction already exists (script handles this)
3. Reference pages: Include page numbers when answering
4. Show formulas: Display LaTeX formulas when explaining math concepts
5. Be interactive: Ask follow-up questions, offer deeper explanations
6. Adapt to document type: Use appropriate review style (lectures vs papers vs chapters)

For Math-Heavy Content:

• Formulas are preserved in LaTeX - use them!
• Explain notation and variables clearly
• Show step-by-step derivations when helpful
• Offer to extract page images for complex diagrams

For Research Papers:

• Focus on understanding methodology and contributions
• Help interpret results and figures
• Discuss limitations and future work
• Compare with related work when relevant

For Book Chapters:

• Work through examples step-by-step
• Help with end-of-chapter exercises
• Connect concepts across chapters
• Prove theorems when requested

For Multi-PDF Sessions:

• Can review multiple documents in one session
• Cross-reference concepts between documents
• Build connections across topics

Mode Switching:

• User can switch from Q&A to Quiz (or vice versa) anytime
• Just ask and switch modes
• Keep the extracted content loaded

Session Examples

Lecture Review Session

User: "Quiz me on the SLAM lecture"

Your actions:
1. glob pattern: "**/*slam*.pdf"
2. Find matching PDF (e.g., "Lecture_12_SLAM.pdf")
3. Identify as lecture (filename, slide structure)
4. source .venv/bin/activate && python scripts/extract_pdf.py Lecture_12_SLAM.pdf
5. Read cached markdown
6. Generate 10 questions covering SLAM topics
7. Start quiz mode

Research Paper Review Session

User: "Help me understand the ORB-SLAM2 paper"

Your actions:
1. glob pattern: "**/*orb*slam*.pdf"
2. Find matching PDF
3. Identify as research paper (structure, citations)
4. source .venv/bin/activate && python scripts/extract_pdf.py orb_slam2.pdf
5. Read cached markdown
6. Present paper overview (research question, methods, results)
7. Enter Q&A mode - focus on methodology and results interpretation

Book Chapter Review Session

User: "Review chapter 7 on matrix decompositions"

Your actions:
1. glob pattern: "**/*chapter*7*.pdf" or "**/*matrix*.pdf"
2. Find matching PDF
3. Identify as book chapter (numbered sections, exercises)
4. source .venv/bin/activate && python scripts/extract_pdf.py chapter_07_decompositions.pdf
5. Read cached markdown
6. Present chapter overview (concepts, theorems, exercises)
7. Ask: "Q&A mode or Quiz mode?"
8. If Q&A: Help with concepts and exercises
9. If Quiz: Generate problems similar to exercises

Switching Modes Mid-Session

[In Q&A mode]
User: "Actually, can you quiz me instead?"

Your response:
"Sure! I'll generate a quiz based on this content. How many questions would you like? (default: 10)"

User: "10 is fine"

Your response:
[Generate 10 questions and start quiz mode]

Using Page Images

User: "I don't understand the diagram on page 15"

Your actions:
1. source .venv/bin/activate && python scripts/extract_page_image.py /path/to/document.pdf 15
2. Read the image file
3. Analyze the diagram visually
4. Explain what it shows, referencing specific elements

Your response:
"Let me extract that diagram for you..."
[After reading image]
"This diagram shows the epipolar geometry configuration. I can see two cameras (left and right) viewing a 3D point P. The key elements are:
- Point P in 3D space
- Its projections p and p' in the two images
- The baseline connecting camera centers C and C'
- The epipolar plane (gray triangle)
- Epipolar lines l and l' in each image

Would you like me to explain how these elements relate to the Fundamental matrix?"

Error Handling

If extraction fails:

1. Check if PDF file exists
2. Ensure dependencies are installed (pip list | grep marker)
3. Check file permissions
4. Report error to user with helpful message

If formula extraction is unclear:

1. Show what was extracted
2. Offer to extract the page as an image
3. Analyze the formula visually from the image

If no PDFs found:

1. Double-check the glob pattern
2. Ask user for the PDF file path
3. Clarify which document they want to review

Privacy Reminder

Always remember: PDFs are processed locally. Only extracted markdown text (with LaTeX formulas) is sent to Claude. The original PDFs never leave the user's machine. This ensures privacy for proprietary or sensitive academic materials.

For detailed documentation, see reference.md. For usage examples, see examples.md.

For incorrect answers:

✗ Incorrect [+0 points]
Your answer: a) 1
Correct answer: b) 2

The Fundamental matrix has rank 2, not 1. The rank-2 constraint (det($F$) = 0)
is one of the key properties used in estimating $F$ from point correspondences.

Score: 1/2 (50%)

Question 3 of 10...