Web Interface

Learn how to use the Aparture web interface for interactive paper analysis.

Overview

The web interface provides a visual, step-by-step workflow for configuring and running paper analysis. It's ideal for:

• First-time setup and exploration
• Testing different configurations
• One-off analyses
• Visual progress monitoring

For automated daily runs, see CLI Automation.

Authentication

When you first visit the application, you'll be prompted for a password.

ACCESS_PASSWORD=your-secure-password-here

This password is set in your .env.local file and protects access to the application and your API keys.

Never share your access password

Anyone with this password can use your API keys and run analyses at your expense.

Configuration Panel

The left sidebar contains all configuration options:

arXiv Categories

Select which arXiv categories to monitor:

1. Expand sections - Click on major categories (cs, stat, astro-ph, etc.)
2. Select subcategories - Check individual boxes or use "Add All"
3. Review selection - Selected categories appear in the summary

Popular selections:

• Machine Learning: cs.LG, cs.AI, stat.ML
• Astrophysics: astro-ph.CO, astro-ph.GA, astro-ph.SR
• Statistics: stat.ML, stat.ME, stat.AP

Research Criteria

Define your research interests in natural language. This prompt is used to score paper relevance.

Example:

I am interested in:
- Deep learning applications in astrophysics
- Bayesian inference methods for time-series data
- Novel neural network architectures
- Interpretable machine learning

Tips:

• Be specific about techniques and applications
• Mention domains and methodologies
• Include both broad interests and specific topics
• Keep it under 200 words for best results

Model Selection

Choose AI models for each processing stage:

Quick Filter (Stage 1)

• Recommended: Claude Haiku 4.5 (fast, cheap)
• Alternative: Gemini Flash-Lite (similar performance)

Abstract Scoring (Stage 2)

• Recommended: Claude Sonnet 4.5 (balanced)
• Alternative: ChatGPT (o1/o3-class), Gemini Pro

PDF Analysis (Stage 3)

• Recommended: Claude Opus 4.1 (best quality)
• Alternative: ChatGPT (o1/o3-class) (good, but less vision capability)

NotebookLM Document

• Fixed: Claude Opus 4.1 (required for quality)

See Model Selection for detailed comparisons.

Batch & Threshold Settings

Batch Size

• Controls how many papers are processed per API request
• Recommended: 10-20 for balance
• Higher = fewer API calls but longer individual requests
• Lower = more API calls but better parallelization

Score Thresholds

• Quick Filter: MAYBE/YES threshold (if enabled)
• Abstract Score: Minimum score for PDF analysis (default: 5.0)
• Post-processing: Minimum adjusted score (if enabled)

PDF Analysis Limit

• Maximum number of papers to analyze deeply
• Recommended: 20-30 for daily runs
• Higher = more comprehensive but longer runtime

Optional Stages

Quick Filter (Stage 1)

• Fast YES/NO/MAYBE filtering before detailed scoring
• Reduces API costs by 50-70% on irrelevant papers
• Enable if: You have broad categories or high volume

Post-Processing

• Re-scores papers using comparative analysis
• Improves consistency across batches
• Enable if: You notice score inconsistencies

NotebookLM Document

• Generates podcast-optimized document
• Required for podcast generation
• Duration: 5-30 minutes (configure in settings)

Running Analysis

Start Analysis

1. Review configuration - Check all settings are correct
2. Click "Start Analysis" - Workflow begins
3. Monitor progress - Watch real-time updates

Progress Stages

Fetching Papers

• Downloads metadata from arXiv
• Duration: 1-5 minutes
• Shows: Categories, date range, paper count

Quick Filter (if enabled)

• Fast YES/NO/MAYBE classification
• Duration: 2-5 minutes
• Shows: YES/MAYBE/NO counts

Scoring Abstracts

• Detailed relevance scoring (0-10 scale)
• Duration: 10-30 minutes
• Shows: Progress, score distribution

Post-Processing (if enabled)

• Comparative re-scoring
• Duration: 5-15 minutes
• Shows: Score adjustments

Analyzing PDFs

• Deep analysis of top papers
• Duration: 15-45 minutes
• Shows: Current paper, progress

Generating NotebookLM Document (if enabled)

• Creates podcast-optimized summary
• Duration: 1-3 minutes
• Shows: Processing status

Monitoring

During analysis, you'll see:

• Current stage - Which processing step is active
• Progress bars - Completion percentage
• Paper counts - Papers processed/remaining
• Estimated time - Based on current speed
• Live results - Top papers updating in real-time

Can I close my browser?

No, the web interface requires an active connection. For unattended runs, use CLI automation.

Results Panel

As analysis progresses, results appear in the main panel:

Paper Cards

Each paper shows:

• Title - Linked to arXiv abstract
• Authors - Primary authors listed
• Score - Relevance score (0-10)
• Abstract - With relevance justification
• PDF Analysis - Deep summary (if analyzed)

Card colors:

• Green border - High relevance (8-10)
• Yellow border - Moderate relevance (5-7)
• No border - Lower relevance (<5)

Sorting & Filtering

• Sort by score - Highest relevance first (default)
• Sort by date - Newest papers first
• Filter by stage - Show only PDF-analyzed papers
• Search - Find papers by keyword

Downloading Reports

When analysis completes:

Analysis Report

Click "Download Report" to get a comprehensive markdown file:

Includes:

• Executive summary
• Top papers with scores and justifications
• Full PDF analyses
• Configuration details
• Timing and cost information

Filename: YYYY-MM-DD_arxiv_analysis_XXmin.md

NotebookLM Document

If generated, click "Download NotebookLM Document":

Includes:

• Structured summaries optimized for audio
• Paper highlights and connections
• Thematic organization
• Ready for podcast generation

Filename: YYYY-MM-DD_notebooklm_XXmin.md

Want podcasts?

Upload the NotebookLM document to notebooklm.google.com and click "Generate" in the Audio Overview section. Or use CLI automation to automate this.

Keyboard Shortcuts

• Ctrl/Cmd + K - Focus search
• Ctrl/Cmd + D - Download report
• Esc - Close modals

Troubleshooting

Analysis Won't Start

Check:

• At least one category selected
• All required environment variables set
• Browser console for errors (F12)

Progress Stuck

Try:

• Check browser console for errors (F12)
• Verify API keys are valid and have available credits
• Check API rate limits in provider dashboards
• Refresh the page if interface becomes unresponsive (you'll need to restart analysis)

Papers Not Loading

Causes:

• No papers published today in selected categories
• arXiv API temporarily unavailable
• Network connectivity issues

Solution:

• Try different categories
• Wait a few minutes and retry
• Check arXiv status page

High API Costs

Reduce costs:

• Enable Quick Filter to pre-screen papers
• Lower batch sizes for better control
• Use cheaper models (Haiku instead of Opus)
• Reduce PDF analysis limit

Next Steps

• Learn about CLI automation →
• Understand multi-stage analysis →
• Test your setup →