This is a neat product, and I plan on trying out some of the recommendations for sunscreen.
During my journey using the app there were a few things I noticed
1) It seems like the intermediate page is generating text from the LLM as well, which makes the whole process quite slow on my machine. It took maybe 10 seconds before the loader finished displaying the text. If I try and perform the same query again on the same browser, the results are somewhat quicker, maybe 700-800ms of wait time, but this still seems too slow. Once I ran the query five or so times, it was as quick as the demo queries on the front page.
2) Consistent results: If I use the same query on separate browsers, I'm given different products as the "Top Recommended Product", which seems odd. I know LLMs are stochastic, but the feed starting with the "Top Recommended Product" probably shouldn't have stochasticity. This problem opens up some interesting ML cans of worms, but I believe these issues could be overcome.
3) Another issue was if I wanted to scroll in the left column while the right column was still loading, the scrolling was very janky. This was an issue on firefox, but it took quite a long time for the app to be functional (> 10s)
4) Perhaps you could move the search bar and the logo to the top, so the logo is on the top left corner and the search bar takes space to the right of it. This way there aren't overlapping elements, I'm sure there's some annoying edge cases there which would frustrate users
5) For negative ingredients (and maybe any of the ingredients) it would be nice if you kept track of an ingredient database with references. I want to know why some ingredient is bad for my skin, and what I could expect.
6) If a product has many distributors, my first through was the arrow scrolling through products was a slider for the distributor list. I wonder if there's a nice way to differentiate the arrow further, so its functionality is more apparent.
Anyway, this is an excellent proof of concept, I'm excited to see how this product develops.
As for the performance issues, we're looking into several things that could speed things up
- Fine-tuning a small LLM for the results on the intermediate page and deploying on a provider with higher throughput and time to first token
- Admittedly, there there are quite a few SQL query / index optimizations we need to make on the backend, along with making parts of our pipeline async
- The frontend itself is also not very performant right now, but we're working on it.
We cache previous calls to the API, so that's why the demo queries or queries others have tried before you are faster. I'll ship a change that makes the results more consistent but not fully consistent later today.
As for the ingredients, citing sources is definitely a next step. In the meantime, I recommend looking up the ingredients that catch your eye on a place like EWG Skin Deep if it's a huge concern for you (I used to do this to make sure my ingredients weren't comedogenic for acne).
Great point about the distributor list UI, we'll think about a better way to show it!
During my journey using the app there were a few things I noticed
1) It seems like the intermediate page is generating text from the LLM as well, which makes the whole process quite slow on my machine. It took maybe 10 seconds before the loader finished displaying the text. If I try and perform the same query again on the same browser, the results are somewhat quicker, maybe 700-800ms of wait time, but this still seems too slow. Once I ran the query five or so times, it was as quick as the demo queries on the front page.
2) Consistent results: If I use the same query on separate browsers, I'm given different products as the "Top Recommended Product", which seems odd. I know LLMs are stochastic, but the feed starting with the "Top Recommended Product" probably shouldn't have stochasticity. This problem opens up some interesting ML cans of worms, but I believe these issues could be overcome.
3) Another issue was if I wanted to scroll in the left column while the right column was still loading, the scrolling was very janky. This was an issue on firefox, but it took quite a long time for the app to be functional (> 10s)
4) Perhaps you could move the search bar and the logo to the top, so the logo is on the top left corner and the search bar takes space to the right of it. This way there aren't overlapping elements, I'm sure there's some annoying edge cases there which would frustrate users
5) For negative ingredients (and maybe any of the ingredients) it would be nice if you kept track of an ingredient database with references. I want to know why some ingredient is bad for my skin, and what I could expect.
6) If a product has many distributors, my first through was the arrow scrolling through products was a slider for the distributor list. I wonder if there's a nice way to differentiate the arrow further, so its functionality is more apparent.
Anyway, this is an excellent proof of concept, I'm excited to see how this product develops.