This is a submission for the Cloudflare AI Challenge.
What I Built
Tafy is an intuitive meal planner that takes all of the pain out of planning your meals for the week by generating recipes.
Demo
You can demo the application here --> https://tafy.recipes
My Code
You can peruse at your leisure here: tafy-demo
Journey
This project has been something that I've been stewing on for years (pun intended)! I decided to pursue it as a business, then shortly after ChatGPT came out and I thought I was done for ... but then I realized that it would just make it that much better!
I decided to do this as soon as I heard about the challenge (about 5 days ago!) so I got to work, and decided to beg my family for forgiveness later - sorry, honey!
In my work, I've been doing freelance with LLMs and generative AI and so I've been following the space heavily, and learned recently about CopilotKit in an article on dev.to. So I had to try it!
With that, I had to hack it a bit to get the tooling to work with Cloudflare since a lot of text generation models don't support calling tools. Luckily, @hf/nousresearch/hermes-2-pro-mistral-7b hosted in Cloudflare Workers AI was able to rescue me, but the model did break my search integration.
I also decided to go with the @cf/bytedance/stable-diffusion-xl model for generating images, and so far I've been really impressed!
Finally, I'm using OpenAI's text-to-speech model for verbal descriptions of the meal plans. Though I was hoping to turn this into a means of getting a recipe read off to you while you make it, I actually like the smooth-jazz DJ AI voice-over while I'm looking at the sweet pics of the food that it comes up with!
I'm proud of what I've been able to accomplish what I did in such a short time, but I couldn't have done it without the help of folks in chat - thank you!!!
Check out my README on GitHub for a list of things that I'll be working on next, and maybe someday, if everything works out, I'll make (this website)[https://tafy.io] live!
Multiple Models and/or Triple Task Types
I used multiple models and tasks:
Models
@hf/nousresearch/hermes-2-pro-mistral-7b@cf/bytedance/stable-diffusion-xl- OpenAI text-to-speech
Tasks
I'm not sure what constitutes as a task, so I'll take a shot here...
[x] The LLM calls tools which equate to "research," each a text generation task in its own right: search, curate, write, critique, revise.
[x] The recipe's summary is sent to OpenAI's text-to-speech API.
[x] Cloudflare Worker - text to image conversion.
[x] Web search (partial, this works with GPT-4, but not Hermes because I can't get it to output JSON).
Things I didn't get to, but may be partially setup:
[] Vectorize store with recipes deposited and used to supplement.
[] Preferences stored on disk, in a DB, or in object storage.
Top comments (0)