Skip to main content

Recipe Parser: Better design? Or design better?

· 6 min read

I never found a solution for storing favorite recipes that wasn't very manual or required signing up for a service. As I explore the world of software development and solutions-as-code, I chose this as a problem statement for a DIY approach. What I really should do is find a cheap printer and have a physical binder full of recipes like they did in the olden days, but that's no fun...

The requirements for this project were simple, or so I thought: Save recipes from the web in a readily accessible and usable format. I didn't realize until later that I had many more requirements I had not discovered. In the end, valuable lessons were learning about knowing your requirements and... just knowing more?

Timeline​

What have other people done?​

I saw many recommendations for Copy Me That, Paprika App, and plain-old recipe cards, but nothing that satisfied. Many of my requirements were formed at this phase, though they lacked firm statements.

Requirements Discovered
  • Didn't want to pay for a service
  • Free services hosted by others can be liability in privacy and vendor lock-in

Start somewhere​

Since I'm already immersed in the Google ecosystem, and they've taken my data already, Drive was a familiar choice. At the risk of being shoehorned into a less-than-ideal platform, I created a folder for storing recipes as Google Docs and a Google Sheet for acting as the table of contents. At this point, I considered manually copying and formatting the recipes as a cop-out answer, but I forged ahead into uncharted territory.

Recipes/ 
├── Table-Of-Contents
└── RecipeDocs/ 
	├── Recipe1 
	└── Recipe2

Throw AI at it​

I knew that any AI chat bot would be able to browse to a recipe article and extract the ingredients list and instructions. At the very least, an LLM would be effective at extracting content from raw HTML. I installed Ollama for the first time and played around with gemma3:1b and deepseekr1:1.5b models. While the novelty of running AI on my own computer was exciting, I wasn't any closer to a final product. With this newfound knowledge, and the smallest bit of experience, I was able to form more requirements.

Requirements Discovered
  • Can't be local to just my desktop
  • Needs to be accessible from anywhere, not just my home network
  • Avoid the security burden and not be publicly accessible

If not on-prem, then cloud​

Cloudflare was the first stop, being the cloud provider that I had the most familiarity with. Even though Cloudflare's compute offerings pale in comparison to the other titans in the industry, finding a solution would mean one less account I had to sign up for.

I came across several offerings that would help me in my search, the first being Email Workers. A serverless function triggered on the receipt of an email. The flow of the solution was beginning to form in my head:

  1. Send email with recipe page URL to destination address
  2. Parse email to obtain URL
  3. Fetch HTML content
  4. Prompt LLM to extract recipe info
  5. Output to Google Docs

The second offering I discovered was Workers AI. A serverless integration for LLMs with generous free tiers on a known platform. I created a prompt and received a response from the LLM - nice and easy.

The third offering was Browser Rendering, to run headless Chrome for automation and web scraping. Looking back, I should have stopped at the /browser-rendering/content API and sent the raw HTML into an LLM prompt but I was more concerned with finding the "best" solution. I spent the majority of my time working with an integration called Stagehand, an AI powered browser automation library. After code incompatibilities, version conflicts, and overall sub-par performance, the email --> LLM flow achieved a usable output.

Screenshot of terminal

Requirements Discovered
  • Needs to be maintainable end-to-end by me
  • Simple is better
  • Avoid hundreds of dependencies (npm packages)

Time to integrate with Google​

To fulfil the end goal, the output should be saved to a Google Doc. At some point while looking at the Node.js Drive SDK I asked myself "If I have to integrate with Google anyway, why not just use Google for everything?" This was the beginning of the end, as I bounced around between Google Cloud, Google Workspace and eventually found something I had never heard of before, Google Apps Scripts. There were some limitations, but a new code flow formed in my head:

  1. Send email with recipe page URL to personal email
  2. Flag email with "Automation" label
  3. Parse email to obtain URL
  4. Fetch HTML content
  5. Prompt LLM to extract recipe info
  6. Output to Google Docs

I left all the progress made with the previous solution, somewhat disappointingly, with the hope I could make something that better suits my needs. Within a few hours I had a fully working product. Project available on my GitHub: https://github.com/lostinstarvation/google-recipe-parser

There are downsides to the current approach:

  1. No event based trigger, so the execution of the script has to be scheduled
  2. Dependent on email rules to correctly identify emails
  3. Runs under user account context

Lessons learned​

My requirements looked like this - a stark contrast from the simple statement in the beginning:

  1. Have full control over solution to add or remove features
  2. Minimize third party integrations and dependencies
  3. Free, without paying in privacy
  4. Recipes should be exportable or saved in a common format
  5. Accessible from desktop or mobile while away or at home
  6. Easy to use for multiple people

So, to answer the question, do I need to be better at designing solutions or know how to design solutions better?

  • Better at designing solutions: Knowing the right tools for the job comes with time, as you can't know everything from the start
  • Design solutions better: One improvement, out of many I'm sure, is to know your requirements