NotebookLM Integration

Understanding Aparture's NotebookLM integration and podcast generation.

What is NotebookLM?

NotebookLM is Google's AI-powered research assistant that can:

• Synthesize information across multiple documents
• Generate conversational audio overviews (podcasts)
• Create natural dialogue between two AI hosts
• Produce high-quality audio summaries

Aparture integrates with NotebookLM to transform your daily paper analysis into engaging podcasts.

The Workflow

Manual Workflow (Web Interface)

1. Run Aparture analysis - Generate report
2. Generate NotebookLM document - Click "Generate NotebookLM Document"
3. Download document - Get .md file
4. Upload to NotebookLM - Visit notebooklm.google.com, create notebook, upload file
5. Generate audio - Click "Generate" in Audio Overview section
6. Download podcast - Wait 10-20 minutes, download .m4a file

Automated Workflow (CLI)

npm run analyze

Complete automation:

• Generates analysis report
• Creates NotebookLM document
• Uploads to NotebookLM automatically
• Generates podcast
• Downloads audio file

See CLI Automation for details.

First-time setup

CLI automation requires interactive Google login on first run. Subsequent runs are fully automated.

NotebookLM Document Structure

The NotebookLM document is specifically structured for audio generation:

Overview Section

# Research Highlights: Computer Science (cs.LG, cs.AI)

Today's analysis examined 47 papers in machine learning and AI.
Several exciting developments emerged across transformer efficiency,
Bayesian methods, and interpretability research.

Purpose: Set context and frame the discussion

Characteristics:

• Conversational tone
• High-level summary
• Engaging opening

Major Themes

## Major Themes

### Transformer Efficiency

Multiple papers tackle the computational challenges of transformers.
Smith et al. introduce "Sparse Hierarchical Attention" which reduces
complexity from O(n²) to O(n log n) while maintaining accuracy...

### Bayesian Deep Learning

There's renewed interest in Bayesian approaches. Jones et al. present
a scalable method for uncertainty estimation in neural networks...

Purpose: Organize papers into coherent themes

Characteristics:

• Thematic grouping
• Clear section headers
• Connected narrative

Deep Dives

## Deep Dive: Sparse Hierarchical Attention

This paper by Smith et al. addresses a fundamental challenge in AI:
transformers are computationally expensive on long sequences.

**The Problem:** Standard attention requires O(n²) operations, making
it impractical for documents beyond a few thousand tokens.

**The Approach:** The authors use hierarchical clustering to create
sparse attention patterns that approximate full attention.

**The Results:** Three times faster training with minimal loss in
accuracy. This opens doors to processing much longer documents.

**Why It Matters:** Could enable new applications in document
understanding, code analysis, and scientific literature review.

Purpose: Provide detailed analysis of top papers

Characteristics:

• Problem-Solution-Impact structure
• Concrete details
• Context and implications
• Audio-friendly narrative

Connections and Insights

## Connecting the Dots

Several papers share common threads:

- Both Smith et al. and Chen et al. use hierarchical structures to
  reduce computational complexity
- Jones et al.'s Bayesian methods could be applied to the attention
  mechanisms in Lee et al.'s work
- The interpretability focus in Brown et al. aligns with growing
  concerns about black-box models

These connections suggest a broader shift toward efficient,
interpretable, and uncertainty-aware AI systems.

Purpose: Synthesize insights across papers

Characteristics:

• Cross-paper connections
• Broader implications
• Forward-looking perspective

Podcast Customization

Duration Settings

Configure podcast length in Aparture settings:

• 5 minutes - Quick highlights
• 10 minutes - Standard summary
• 15 minutes - Detailed overview (default)
• 20 minutes - Comprehensive discussion
• 25 minutes - Extended analysis
• 30 minutes - Deep dive

Recommendation: 15-20 minutes for daily listening

Custom Prompts

Aparture uses custom prompts to guide podcast generation. Prompts are defined in NOTEBOOKLM_PROMPTS.md:

5-minute prompt (excerpt):

Focus on the top 3-5 most significant papers only. Keep technical
depth appropriate for audio but maintain accuracy. Each paper should
get 30-60 seconds of discussion maximum...

15-minute prompt (excerpt):

Provide a balanced overview covering:
- 5-7 top papers with moderate technical depth
- Key methodological approaches
- Implications and connections between papers
Target 2-3 minutes per major paper...

30-minute prompt (excerpt):

Create a comprehensive discussion covering:
- Detailed analysis of top papers
- Technical methodology discussion
- Broader context and related work
- Connections across papers
Aim for 3-5 minutes per major paper...

Customizing Prompts

Edit NOTEBOOKLM_PROMPTS.md to customize podcast style:

Example customizations:

• Add domain-specific context
• Adjust technical depth
• Emphasize certain aspects (methodology, results, implications)
• Change conversation style
• Add humor or personality

Iterative refinement

Generate a few podcasts with default prompts, then adjust based on what you want more/less of.

Audio Quality

NotebookLM Audio Features

Voice characteristics:

• Two AI hosts (one male, one female by default)
• Natural conversation flow
• Appropriate intonation and emphasis
• Technical terms pronounced correctly (usually)

Audio format:

• File format: M4A (AAC encoding)
• Bitrate: ~128 kbps
• Sample rate: 44.1 kHz
• File size: ~1-3 MB per minute

Common Issues

Mispronunciations:

• Novel technical terms
• Author names (especially non-English)
• Acronyms

Solution: Edit NotebookLM document to include phonetic hints or expand acronyms

Awkward transitions:

• Abrupt topic changes
• Repetitive phrases

Solution: Improve document structure with better thematic organization

Too technical or too simple:

• Doesn't match your background

Solution: Adjust custom prompts to target your knowledge level

Use Cases

Daily Commute

Scenario: 30-minute drive, want to stay current

Configuration:

• 20-25 minute podcast
• Balanced technical depth
• Focus on top papers

Workflow:

1. Run npm run analyze before bed
2. Podcast ready in the morning
3. Listen during commute
4. Read full papers that interest you

Research Group Meetings

Scenario: Weekly paper discussions

Configuration:

• 15-20 minute podcast
• Higher technical depth
• Include methodology details

Workflow:

1. Weekly analysis run
2. Share podcast with group
3. Everyone listens before meeting
4. Discussion focuses on implications

Interdisciplinary Exploration

Scenario: Discovering connections across fields

Configuration:

• 25-30 minute podcast
• Multiple diverse categories
• Emphasize connections and themes

Workflow:

1. Select broad category set
2. Generate comprehensive podcast
3. Listen for unexpected connections
4. Deep-dive into promising areas

Literature Review Preparation

Scenario: Starting new research direction

Configuration:

• 30-minute podcast
• Deep technical detail
• Include future directions

Workflow:

1. Run analysis on target categories
2. Generate detailed podcast
3. Listen multiple times
4. Read full PDFs of key papers
5. Repeat weekly to build knowledge

Automation Details

CLI Automation Process

When you run npm run analyze:

1. Analysis completes - Report generated
2. NotebookLM document created - Structured for audio
3. Browser automation starts
   • Opens Chrome with persistent profile
   • Navigates to notebooklm.google.com
   • Authenticates (first run only)
4. File upload
   • Creates new notebook
   • Uploads analysis report
   • Uploads NotebookLM document
5. Podcast generation
   • Opens customization menu
   • Enters duration-specific prompt
   • Clicks "Generate"
6. Wait for completion (10-20 minutes)
   • Polls every 30 seconds
   • Refreshes page to check status
7. Download audio
   • Finds three-dot menu
   • Clicks Download
   • Saves to reports/
8. Cleanup
   • Takes screenshots
   • Closes browser
   • Reports success

Authentication Caching

First run:

• Browser opens to Google login
• You manually sign in and grant permissions
• Session cached in temp/notebooklm-profile/

Subsequent runs:

• Uses cached session
• No manual interaction needed
• Fully automated

Session expiration:

• Google sessions last ~weeks to months
• If expired, you'll need to re-authenticate
• CLI will detect and prompt you

Error Handling

Upload failures:

• Retries up to 3 times
• Takes screenshots for debugging
• Detailed error messages

Generation timeouts:

• 30-minute default timeout
• Configurable in code
• Logs elapsed time regularly

Download failures:

• Retries menu navigation
• Screenshots current state
• Suggests manual download if failed

Cost Considerations

NotebookLM Costs

Good news: NotebookLM is currently free to use.

Google provides:

• Unlimited notebooks
• Unlimited source uploads
• Unlimited audio generation
• No credit card required

Limitations:

• 50 sources per notebook (plenty for daily use)
• Rate limiting (rarely hit with daily runs)
• Subject to change (Google may introduce pricing)

Aparture Costs

NotebookLM document generation:

• Uses Claude Opus 4.1 (fixed)
• Typical cost: $0.20-0.50 per document
• Depends on paper count and detail level

Included in daily costs:

• Balanced config: $3.95/day includes NotebookLM document
• Budget config: $2.05/day includes NotebookLM document

Comparison to Alternatives

Manual Reading

Pros of podcasts:

• Multitask-friendly
• Lower cognitive load
• Good for initial filtering
• Identifies papers for deep reading

Cons of podcasts:

• Less detailed than reading
• Can't skim/skip
• No direct access to figures
• May miss nuances

Best practice: Use podcasts for discovery, read PDFs for deep understanding

Other Audio Services

NotebookLM vs. podcast apps:

• NotebookLM: Custom content, your specific interests
• Podcast apps: General audience, broader topics
• NotebookLM: Free, unlimited
• Podcast apps: Often paid, limited selection

NotebookLM vs. text-to-speech:

• NotebookLM: Conversational, synthesized
• TTS: Literal reading of papers
• NotebookLM: Engaging, contextual
• TTS: Dry, hard to follow

Next Steps

• Set up CLI automation →
• Learn about reports and outputs →
• Compare different models →