Skip to content

Content Extraction

Configure templates, providers, and costs for AI-powered content extraction.

How It Works

Inkwell uses AI to extract structured information from podcast transcripts:

Transcript → Template Selection → AI Extraction → Markdown Output

Each template defines what to extract (quotes, summary, concepts, etc.) and produces a separate markdown file.

Available Templates

Template Description Output
summary Episode overview and key takeaways summary.md
quotes Notable quotes with speaker attribution quotes.md
key-concepts Main ideas and concepts discussed key-concepts.md
tools-mentioned Software, apps, products mentioned tools-mentioned.md
books-mentioned Books and resources referenced books-mentioned.md

Template Selection

Automatic (Default)

Templates are auto-selected based on episode category:

Category Templates
tech summary, quotes, tools-mentioned
business summary, quotes, key-concepts
interview summary, quotes

Manual Selection

Override with --templates:

# Specific templates
inkwell fetch URL --templates summary,quotes

# All templates
inkwell fetch URL --templates summary,quotes,key-concepts,tools-mentioned,books-mentioned

Category Override

Force a category to change auto-selection:

inkwell fetch URL --category tech

Providers

Inkwell supports two AI providers for extraction:

Gemini (Default)

  • Cost: ~$0.003 per template
  • Quality: Good for most use cases
  • Speed: Fast

Claude

  • Cost: ~$0.12 per template (40x more expensive)
  • Quality: Best for precision tasks
  • Speed: Moderate

Smart Selection (Default)

Inkwell automatically chooses based on template:

  • Gemini: summary, key-concepts, tools-mentioned
  • Claude: quotes, books-mentioned (precision matters)

Force Provider

# Use Gemini for everything (lowest cost)
inkwell fetch URL --provider gemini

# Use Claude for everything (highest quality)
inkwell fetch URL --provider claude

Costs

Cost Estimation

Check costs before processing:

inkwell fetch URL --dry-run

Output: 

Estimated cost: $0.0090
  • summary (gemini): $0.0030
  • quotes (claude): $0.0045
  • key-concepts (gemini): $0.0015

Typical Costs

Episode Size Templates Provider Cost
30 min (~5k words) 3 Gemini ~$0.003
30 min (~5k words) 3 Claude ~$0.045
120 min (~20k words) 5 Gemini ~$0.012
120 min (~20k words) 5 Claude ~$0.180

Track Spending

inkwell costs

Output: 

┌ Overall ─────────────────────┐
│ Total Operations:  15         │
│ Total Cost:        $0.0825    │
│                               │
│ By Provider:                  │
│   gemini    $0.0525           │
│   claude    $0.0300           │
│                               │
│ By Operation:                 │
│   extraction    $0.0825       │
└───────────────────────────────┘

Caching

Extractions are cached to save time and money.

How It Works

  • Cache key includes transcript hash, template name/version, provider, model, prompt hash, output schema version, and cache format version
  • Default cache duration: 30 days
  • Cache location: ~/.cache/inkwell/extractions/

Cache Behavior

Scenario Result
Same transcript, same template Cache hit ($0)
Same transcript, different template New extraction
Template version updated Cache invalidated
Prompt, provider, model, or output schema changed Cache invalidated

See Cache Behavior for transcript, extraction, and media/audio cache details.

Skip Cache

Force fresh extraction:

inkwell fetch URL --skip-cache

Concurrency

Templates are extracted in parallel for faster processing:

  • Default: 5 concurrent extractions
  • ~5x speedup compared to sequential

Template Versioning

Templates include version numbers:

# templates/summary.yaml
name: summary
version: 2  # Incremented when prompt changes
expected_format: text

When a template is updated:

  1. New version number invalidates cache
  2. Next extraction uses updated prompt
  3. Old cached results are ignored

Cost Optimization Tips

  1. Use YouTube transcripts - They're free
  2. Default to Gemini - 40x cheaper than Claude
  3. Select fewer templates - Only extract what you need
  4. Leverage caching - Don't re-process unnecessarily
  5. Check with --dry-run - Know costs before committing

Next Steps