Cooker AI: generate meals using AI

Coming up with ideas for cooking sucks? Well... don't.

Introduction

Cooker AI is a website that helps you with this boring task: find ideas for cooking. Finding meal ideas for up to 14 times a week is complicated. That’s 728 ideas a year! Easy solution: let AI generate meals for you.

This website has a bunch of interesting features:

• Plan meals ahead with a calendar-like interface
• Generate ideas based on parameters (diet, allergies, duration…)
• Swap meals if needed
• Regenerate meals if you don’t like them
• Share meal plans with other users
• Favorite meals you liked

For this project, I was the lead fullstack developer. This means I was in charge of the GitHub repository, project management, CI/CD and making code reviews. I also worked a lot on the UX design mockup.

We used the following technologies:

• Ruby on Rails
• Open AI API
• Tailwind CSS
• Stimulus
• Sidekiq
• Cloudinary
• Heroku

Check it out here!

Purpose and Goal

At Le Wagon, the bootcamp ends with a 2-weeks project, that we have to demo the very last day (You can see the video there btw). So I pitched this app during the bootcamp (I already had some related ideas for quite some time) and we built this app with Maria Skrabalekova and Chuma Mukala!

The goal of the project was to reinvest all the knowledge from the previous 2 months into a great (working) app. It had to include challenging features if possible, that’s why we chose AI.

Before actually coding, we spent some time making some UX Research on Figma, as well as defining our user stories and database schema.

Spotlight: AI generation

There are many interesting things in this project (background jobs, websockets) but let’s focus on the main technical point of this project: AI generation. We use 2 models from Open AI:

• gpt-3.5-turbo to generate the meals’ data
• dall-e (256x256 res) to generate meal images (yeah even those are AI-generated)

We faced a few issues though:

Inconsistent results

To actually exploit the data returned by Chat GPT, we needed results in JSON format. However, we can’t force it to return JSON so results could be quite inconsistent.

To improve them, we worked on 2 main topics:

1. A good prompt: here is how it looks like (optimized by Chat GPT itself!):

Please provide meal ideas that meet the following criteria:
- All meals must be returned as JSON objects in the following format:
 {
  "date": "yyyy-mm-dd",
  "dish": "name of the meal (e.g. dinner)",
  "name": "name of the dish",
  "ingredients": ["list of ingredients"],
  "steps": ["list of cooking steps"],
  "prep_time": "time needed to prepare the meal"
 }
- I need ideas for #{attrs[:dishes]} for #{attrs[:days]} day#{attrs[:days].to_i == 1 ? "" : "s"}
- The "date" field should be formatted as "yyyy-mm-dd" and the first date should be #{attrs[:date]}.
- Please consider my allergies to #{attrs[:allergies]}.
- I really enjoy #{attrs[:cuisine]} cuisine.
- I follow a #{attrs[:diet]} diet.
- Only return an array of JSON objects (even if there is only one meal idea).
- No sentences, only JSON.
- Any other format will result in a rejected request.

2. Parsing: Chat GPT often adds comments before (and even after) the actual content such as:

Here are some meal ideas:

[
    {
        "date": "..."
    }
]

So we had to parse this a little bit. It might not be the cleanest way but it works:

def self.parse_response(res)
    # Remove all the \n junk + ...
    res = res.gsub(/\s+/, " ").strip.gsub("...", "")
    # Chat GPT always says sorry when he can't give only json
    valid = !res.downcase.include?("sorry")
    # In this case we need some manual parsing
    unless valid
        # clean everything that is before the first [
        _trash_content, *remainder = res.split("[")
        res = "[#{remainder.join("[")}"
        # If there is something after the json, clean everything
        # that is after the last ]
        unless res[-1, 1] == "]"
            *remainder, _trash_content = res.split("]")
            res = "#{remainder.join("]")}]"
        end
    end
    return JSON.parse(res)
end

Finding recipes using scraping

Because results can be creative with Chat GPT, there were cases where the recipes didn’t exist on websites such as allrecipes. We even tried the Spoonacular food API but we faced the same problem. Solution: even generate the recipes and any other properties with Chat GPT!

Because of the issue above, we also couldn’t find images on websites so the only solution left there was to even generate the images with AI. We found this pretty funny!

It’s slow

This is something we have no control over. Indeed here are the number of requests made to the Open AI API:

So as you can, it depends on the user input. Moreover, it may take more time if it fails (because JSON is not valid) but since we generate in background jobs, we retry up to 3 times. This minimizes the risk of failure but increases the waiting time.

Current status

No more development is planned on this website. We may try to do a bit of marketing to see if it gets interest. In that case we’ll do things but otherwise it’s kind of dead. But feel free to try it yourself!

Lessons Learned

This project taught me several things:

• Interacting with AI through code
• Project management
• CI/CD with Heroku
• I don’t suck too much at coding: doing a lot of code reviews made me realize I have quite some experience now

However, I’m not a big fan of the technologies with used (except Tailwind CSS VS Bootstrap!): even if Ruby on Rails is powerful, I really miss the safety TypeScript brings.