The Road to NG

[intellectronica]

MkFlashcards grew from a script to an applet, and I have now used it for a couple dozen decks, and received some feedback from others, so I feel ready to start working on its next, more mature iteration. Here are some thoughts I have on what's coming:

• Fixed set of models, likely from OpenAI (DONE in Prep for NG #15 )
  • Easier to evolve prompts for a fixed set of models than try to support everything
  • I love Claude, and I would still like to try to use Gemini and maybe some other models, but need to prioritise this well below functionality
  • It may prove beneficial to use a mix of models rather than the same model for all tasks, and standardising on a single provider makes this easier
• Use Pydantic-AI ( Convert to Pydantic AI #13 )
  • BLOCKER, since it’s still pre-alpha, but I’m not in a rush, and they seem to be moving fast
  • It’s going to be great for working with structured outputs
  • It’s integrated with Logfire - great for monitoring and debugging
• Proper server app, with users, sessions, and persistence
  • Need to have an easy way for people to run it locally in single user mode
  • What about API key? Possible solutions:
    • Keep that browser-based and send it over with every request - a bit dumb TBH
    • Record user’s API key on the server (encourage them to have a unique key and limits)
    • Use a central API key and charge users for usage - that would be cool, but how hard is it for a single developer with not too much time to manage a small SaaS?
  • Allow users to set custom instructions
  • Save output of intermediate steps - the content is valuable and can be recycled
  • Managing decks and cards
  • Sharing decks (would be cool to have a library of useful decks)
  • Templating for cards (how? jinja? something simpler?)
  • Allow exporting in multiple useful formats
    • Mochi
    • Anki
    • CSV
    • PDF
    • HTML / Markdown
• Read the text and generate cards in multiple passes. Something like:
  • Read instructions and plan steps
  • Summary
  • Remove unnecessary parts of the text (footnotes, etc…)
  • Chunk (how well can we chunk? doesn’t have to be equal chunks)
  • For each chunk:
    • Decide on topics for cards
    • For each topic:
      • Write questions
      • Answer questions
      • Look for quotes