kidroca / bedtime-story-generator Goto Github PK

Description

Most of the time the generated stories are short and general
Regardless of instructions like "detailed", "deep", "extensive", "use at least a 1000 words" or "each chapter should have at least 4 paragraphs", the generated output is far away from the prompted requirements in terms of length

We're looking for ways to inspire the generation of longer stories

The initial plan is to have a system message set some ground rules and then a user message inspire a story
The idea is to keep it simple and avoid back and forth message like asking for each chapter individually
A story between 1000-2000 words would be ideal

We've tried providing longer user message with more details, though this oftens drives the story around paraphrasing or repeating the user message with a few more words added around
The user message should serve as inspiration and don't need to be retold back at us

Description

It might make more sense to use Firebase rather than the local file system

Unfortunately this would make the project less portable - right now any one can run it locally

As an initial step the naming convention should be improved a bit because it seems to cause problems when the project is hosted on Netlify

• cyrillic characters in file paths may be the culprit
• or white spaces
• or even the : in the ISO date prefix

Ideally when we're listing story files it would be great to see a title or a few words in the file/dir name
Other information like the date can be moved inside the story.json file, though right now we rely on the date a bit for uniqueness

Description

The story generated by the AI would contain a bunch of items:

• title
• chapters
  • title
  • content
  • illustration
• genre

Instead of trying to parse these out, we're asking the AI to output the story in a valid JSON format like

{
  "title": "story title",
  "genre": "story genre", 
  "chapters": [{
      "title": "part title", 
      "content": "chapters content",
      "illustration": "illustration description"
    }]
}

But no matter how much we stress that it should be valid JSON, the generated output would sometime have trivial errors, like using different set of quotes, or leave a trailing comma

The SpeechSynthesisUtterance used to read the story text as audio is not very good for Bulgarian stories

Right now it's setup to read everything in Bulgarian anyway

Tasks

• find something else that read the stories better
• see if the Speech api is any better in English and maybe use it for English while disabling it for BG

I find myself jumping to ChatGPT to ask for guidance or general information to help me improve a certain story
It would be nice to just open a chat with the context of a given story

Description

Some stories turn out better than others, and we haven't found the best parameters yet
It would help further development and help analysis if we store as much information about what caused a certain AI output (the generated story)

This should include the prompts for image generation
For the time being they can be replace the illustration description, because an image prompt would already contain the original description imagePrompt = originalDescription + static style/flavor

Stories generated in English seem to be the best

The thing is, even if we ask the AI to create a story in English and then translate it to other language it looks like the story was directly created in the final language and it's too bland or even plain weird with language errors, weird symbols, especially if temperature is higher than 0.7. Besides we don't really see the original English story that was supposedly translated

Let's create a api/translate where given a Story and a language the story is translated to the specified language and saved separately as a translation (e.g. a new file with metadata like language: BG)

This can be further improved with more parameters like age-range so the story is transformed to a form more appropriate for the target audience

Sometimes we want to edit a chapter - manually or automated to:

• edit some typos
• change character names
• generate more content
• or just rephrase stuff the way we want to

Also we want to be able to re-generate a specific chapter image:

• we want to either re-generate
• update the illustration instructions (prompt) and re-generate
• or just upload an image of our own
• or specify a remote url (that will be downloaded behind the scenes)

It would be great if there's an edit button, that changes the chapter to edit mode where we have some controls to re-generate the story, edit the content manually or provide instructions on how the AI should try and shape it

Also an edit button to change the title or to generate a new one

Description

Add netlify identity handling and allow API access to authorized users only