
I added the image stuff for two parts of Productsbolt. It works for sites utilising on for page use and Shopify websites. I still have to add it for sites that I'm able to acquire with request.js but as far as I can tell, those sites don't seem problematic at all so I'm going to have to investigate that one day. Ah well.

### Has adding images helped

Without tests, I can't prove that. For the last day, I've been editted the prompt without going through tests. The simple rule was, are the pages being added and are in stock. If they are, I'd like to see if they are positive or false positives. This comes pretty easily as you can see the prices and the urls being suspect, so I've leveraged this for now. For the majority of things, it's working fine but there are scenarios which are silly and I need to make sure the prompt knows not to add. Iterating the prompt is pretty simple given I have all the information and functions with endpoint made available. It's simply just adding in a webpageId and shopProductId which allows me to retest individual pages on demand.

### Shouldn't the images help all of the time

I have the feeling images can also be wrong here and there so it's more being used as supplement to the information I was originally giving it and the prompt I'm working with. I have the feeling less false positives are coming up so I'll work on this over and over again for the next day or two to get this sorted.

### Slow but surely

It's slow work because it's LLMs. If I had more powerful hardware, this would be a lot faster so I'm going to have to be a bit smarter on how I'm going to fix this. Given it's LLM work, it's just grunt work getting the prompt to the right place and when it is, hoping it'll stay. I did change the model and it might require a change to a bigger model, but I really really don't want to be moving to 8B over 4B so I'm going to do everything I can to get it to work with 4B and leverage all the information.

### Conclusion

This is all about product accuracy and I'm really getting close now. Yep there's still some false positives but it's getting smaller and smaller to the point, I'll be able to figure out something which will allow me to do a 5 min glance per day to see new products and question if there's a false positive. For now, I'll batter my way through this.


I'm actually not sure if adding images are helping but I sure hope it is!

Alan Reid

Hopefully images are helping


Okay it's been 19 days but it feels like it's been a month! I've put some effort to get back in the groove of constantly coding and to be honest, it's enjoyable but in some sense, conflicting. When I built this website, which I think was 2022, I that's when I really thought of Skyscanner Plus, and I think I coded a lot. Sitting on my sofa from 08:00 to 01:00, just coding away sometimes and getting things made. Tinkering and trying to get things I wanted sorted. It was good to be able to do that and to be honest, I'm probably able to do that again so I'll look into ways to be even more productive with the coding. I'll figure that out.

### So what's conflicting?

I've enjoyed being able to go downstairs at 18:00 and watch my anime sesh. That said, yesterday I did feel like I had burnt myself out on my anime binges and maybe it was time to watch a normal movie, Heartbreak Ridge. Not go be taken too seriously but it was enjoyable to Tom Highway (Clint) do all he could to make sure his men were battle ready. To not die on the battlefield. To be the best person we can be (however conflicting that can sound).

I guess for a few years, it was all I was doing but I have to argue, it made me good at what I was doing. I enjoy being very good at something or having the skill to make things on demand. The conflict comes with the balance of life and myself. After I lost my job at 2024, I tried to balance life a little better. Don't code all the time, don't just be a singular person and try to spread out. It's now 2026 and I have to be honest, I'm not sure if I'm made that way. Looking back, it was either gaming to try to be the very best in the world in vCoD, managing in the many titles I operated with and made sure my players were best equiped as I could to win titles and quite recently, coding/building/developing.

### The logic of life might not apply

It's very possible I'm a very one dimensional person and trying to do many of things and therefore limiting myself the time to actually get very good at something, is problematic and I believe the two years have evidenced that. To be honest I'd love to watch anime all of the time and getting a real mastery of the understanding of the story telling, though let's be honest, most of the stuff is trash so that might be wasted on me. Gaming, honestly I'd love to be way way better but again, I'm not in the sort of position in life anymore that affords me the ability to just game all the time like I was back in 2000s + 2010s. Here's the thing, I like building things and the years coming up to now, I've gotten better. Am I amazing, no. I feel I have so much more to learn and I'm the type of person that things come slow to. I think because of this, I have to accept that putting the extra time and effort, is the way to go and figure a way to make everything else work to it.

### Is it even a good idea with the way the industry is moving

I'm in the belief the skills will always be needed. Knowing how to solve the problem using Claude Code, affords you more time to do more interesting things. Per like trying to setup a vLLM, I'm having to RELY on LLMs to help me set things up. This is not the way to go and I really should have a better understanding of why it's so hard to get vLLM to operate sometimes. However I'm outsourcing that to and LLM and that's not a good idea in the long term. Now if I knew what the problem was or something about it and the outputs were making sense to me, I would be able to be more proactive but I'm being reactive, and I'm one layer in "knowledge debt", where another layer, means I am fully reliant on an LLM. I don't think that's the way it should be.

### So what's the answer

I have the feeling that just going on the grind and doing work constantly is the only way to go about it. It kinda sucks but I mean, it does work and I have the past to showcase that it's mostly true. Annoying though. It's always been the case with games which is what I'm probably using as a reference. If I slack of, I don't do very well, but if I do put the effort in, I'm really up there "though less and less as I got older". In the arena of programming, I think i'll give it a shot again and just nerd it out or put in a lot more hours, rather than think it can be done another day.


I actually thought it was a month but tbh, it's less than 3 weeks

Three weeks going strong


There's only so much prompting you can do until the lack of evidence you can pass through it, allows the prompt to be right more times than wrong. This is problematic whenever I want the prompt to be correct all the time. There's a point where one has to accept, the human can precive issues a lot quicker where all the facts become available.

### Descriptions aren't enough

While I can give the LLM the title of the page, the product name discovered and the description, this is all based upon the shop owner doing the correct thing. From what I can tell, descriptions can be generic. They can be copy and pasted and even though that happens, one can still go onto the webpage and know what they are buying just from the data points one has. While the name, the description and even web url can be generic, you can look at the price and picture to get the rest of the information. While some products can go up in price to the point one thinks, is this the right price, usually the image is the dead give away, and that's where I'm going to work with.

### Why the push back all this time

It's possible the image could actually be problematic for the prompt but it dawned to me, it's only ever going to be a problem in variants and in that case, the variants usually surrender the information I actually need, which thinking about it, I should add in the details.

Point being, in most situations, the webpage image should suffice. This does mean I have to load images again but that's whatever I'll deal with it.

### Method

I'm just going to get the webpage screenshot and work with that. Combined with the description of the page and the that image, the LLM should be smart enough to infer if the image is a pack, a box or something just from general understanding. It should probably be the tipping point.

### Model

Yep that's the problem, I'm going to have to surrender to the idea of changing the model now. I'll still be able to use Qwen 4B but it's going to be a more updated version so this could be a complete break to the application. What I'll do for now is try to get the tests as far as I can but right now, the issue is very fringe and doesn't happen too much and it's easy to recognise and blacklist for now. The lack of details is a big one though.

### Could use price

Tecnically speaking I could look at price also and state that if the product is really far off, it's most likely x or y. It's a bit annoying to think of that because prices of things can go up and down depending on the average. Then again I can always change that with the data I have, so I can just do a quick check myself as I'm already doing to get the average price and then go "well the product should be around this price, if too far away, just say no"


Picture says a thousand words and a lot more. Learnt that the hard way

A picture says a thousand words


For a good wee while now, some sites have many products for a single url and this has always been annoying. To be honest I always said to myself, I'd just use the blacklist and get it sorted that way, but that has been a bad idea. It takes away from actually getting the problem sorted and the solution can sometimes be boring to figure out and code. Sometimes the solution isn't actually about the code but rather the tests to help get to that point, which I think I'll talk about in my next post tomorrow.

Anyway, variants aren't exactly the hardest things to get sorted but it was just annoying to have to get it sorted. It goes against the whole async idea of Productsbolt where I gather as much of the information as possible and pass it on. There probably is a solution and if I can think of it, I will, but right now, when there's more than one variant of a product on a main page, an LLM has to be used to identify which variant to choose. I believe I need state for now.

### Why is state needed?

There's an if block so where a site might not be using Shopify or another technology, my if statement checks if it is a shopify site or not and does something with that. If there's no variant, go to the next step but if there is, figure out what variant there is and then go to the next step. The issue is, we need to go to the next step. This is probably the issue. I've made a function which relies on state immediately. I think there's a solution here.

### Napkin solution

Whenever a product has a variant, I could offload that to the LLM to sort out but then I'm going to have to creat a new pipeline of work. It would require us to determine what the variant is, pass that to the LLM service with information that isn't required for the LLM service itself but for communicating back to the API. From there, we could use a simple event emitter to throw it back where it all started but now ignore the extra variants and just use the variant we chose, thus then it would go through the if statement section of having just one variant, the one we want, and go to the next stage. This is a solution.

### Does it need to be done now?

Honestly, no. It's a pretty fast service and while it might have to wait on other prompts to get done, thus holding up the queue, I heavily doubt it'll be noticeable with the scale I'm currently working with. It might matter in the future so I'll write up an issue to get it done.

### Are we getting closer?

Yes we are. With the links being fixed and variants being looked at, it is now just about the quality of the prompt. There are situations where pages will make it near impossible to know what the page is selling. I'm not going to worry much about that but test on real world situations that Productsbolt gets worked in. As per like fixing the variants and links, I can do stuff which helps make it easier to limit the harder scenarios which shouldn't be coming up anyway.

Really it's just about the quality of the analysis. Can Productsbolt identify correct the product and state of it and if so, is it all of the time. I have the feeling I'm at that point now, and a framework to test this constantly will be required.

### Feedback framework

I'm going to need a way to look at webpages being added and determine if they are correct or not. I'll probably add a `inspected` property on webpages to know if I've inspected the page or not. I'll create a find function which selects a random page which is `inspected === false` and then check constantly. The more this is done, the more changes I'll find a page which is slipped through, to which case I'll add it to the tests and see if all the pages can pass or not. This helps with 2 things.

- Find pages which are legit getting through

This is simple/annoying. Figure out why the prompt is failing and make sure it doesn't in the future. Make sure all other tests pass.

- Behaviour is failing elsewhere

Bit of brainwork needed here. Before answering the first issue above, I need to figure "is the page rightfully being checked for". Per like the links scenario, pages were coming through which could have been filtered out beforehand, which prevents any need for critique. Doesn't take long to figure this out.

This fall under adding tests so I'll add the functionality to a branch I was using to create more tests and this should then allow one to know if the page has been changed for or not and if it has, set to true and then move on.

### Conclusion

Quality is now the question. I want to make sure Productsbolt is doing what I set out to get done and I think it's nearly there. I've went about this in a very silly way. I looked to optimise, get other things done, features that help with after getting products, but I haven't followed exactly what the point of Productsbolt was at it's core, and it's something I've learnt. While everything else helps get Productsbolt useful, per like a car, it's pretty pointless knowing the amount of fuel in the car, if the car doesn't do what it actually needs to do. Get from A/B in a predictable way.

A document showcasing what the general goal of the project is, should be made for every project in the future.


Variants have always been annoying to deal with. This sorts out a big chunk of that issue

Variants fixed! Thank goodness for LLMs


So I was going through making those tests and I got into a situation where the page itself for the product, had information which suggested it was something else, therefore putting me in a tricky position. If the page decsription describes itself as something that it's not, the LLM isn't going to know that. This happens because shops and their staff, copy and paste description information. A human can look at the page and know what that page is about immediately. In some cases, we can't and that difficulity can be the difference of using that shop or not. Funny enough though, this problem shouldn't happen much and it only did because of how I rank links!

### Links were faulty

The prompt used to rank the list was essentially faulty. I give the prompt context about the product and gave it links, but without being specific enough that the links had to accessed seperately rather than one URL entity (weird I know), it was treating the list of coma seperated links, as one thing. This meant if a link wasn't what I wanted but another was, it could conflate it.

There was another problem where the context was being used to essentially add to the url. I had to explain the context was not to be used as addition to the URLs. A bit annoying I had to go to these lengths but I'm assuming this is the limitation of small instruct models. One has to be very specific and direct.

After working on these things, things got a lot better but there's still more work.

### Websites and their APIs

A product can have variations. Currently I'm treating website don't have so if I'm access an API, I just get the price and description and stock status. Unfortunately some website use this functionality rather than creating a new page per product. It's actually quite annoying from a user perspective also who wants to get the information asap or the exact information. However this does allow for the collection of products. To the point, if the API responds with having many of variations, I don't think I'm going to be able to grab the information as is.

#### Leverage LLM

I'll probably have to make a brand new prompt which will look at the API requst for the variants and give the prompt enough information to know how to recognise which of the variants it wants. Yes I can easily do this as a person but in scale, I want this to do it on it's own. So I think it'll just be getting the variants, see if it's just one and then process as normal but if there's more than one, then to use an LLM to determine which of the two should I use. Thank goodness we can extract information automatically this way in a predictable way.

This is pretty important to get done. In this situation, it could be picking the wrong product and information so I'll work on this tomorrow. While there are more tests to get done, the more prompt specific the prompt comes, the harder it'll become to make sure it passes more and more tests. Reducing the possible scenarios will help determine which tests need to be done.

### Conclusion

Happy to have saw the issue with the links being ranked. It'll help with reducing the amount of incorrect pages being labelled. More progress.

On another note, think I need to take it easy with the lifting, bench press wasn't great so I think I need time to rest. Good craic.


Prompt enginnering can be very annoying. Be nice if the LLM talked

A bug when finding new links was found


The sitemap is sorted and a check to see if webpages that are saved on Productsbolt but not on the sitemaps repersenting said pages, was also created. I thought about the logic of having a test which looks at the page and decides, if the page still holds up. I'm going to write some logic in the future which allows for this small problem to get sorted but right now, I could literally just do a webpage reset every so often. It's a very easy fix to solve the problem and computationally, doesn't really add much more. I think i'm more interested to get things running and so, I'll setup something that'll do a CRON job once a week to delete all the webpages and then search for them again. That way, it'll be on top of it. Technically could do it once a day for the laugh.

### What is there now to do?

I think the big one now is making a way to test pages and go "is this actually doing the thing I want" because while I can tolerate issues, if I were to actually release this, I'm going to have to make sure I'm on the ball here. Is this product actually the cheapest, is this even the product in question, am I going to have to intervene constantly? The point of using an LLM, was to try to remove myself and a person from having to come up with those interventions. I'm very pleased with adding the tests a while back to see if the prompts work or not. Lots of fixes are working but I've added more and more sad paths than happy paths so I'll have to work on that again. Thankfully I save most fo the data I need for tests so it's quite straight forward to test.

### Have I tried a stronger model

OpenAI's OSS20B was very interesting to me but given how big it was, I determined that it wouldn't be viable longterm with the setup I was working with. I also noticed that reasoning models weren't deterministic enough. In the future, I'll look into this and see if other models will do a better job but I haven't needed to move from Qwen 4B:2507 and to be honest, being able to test the model constantly against real world situations, helps a ton.

### Conclusion

That's really it. It's making sure the products are as what they said they are. If I get that, this is ready. Everything else has been to get to this point but the point of Productsbotl was making sure one was getting the cheapest product somewhere which was actually available. If I'm classifying a product differently than what it is, it's a complete failure and so, today I'll figure a way to go about this. I can't believe it but I think I'm nearly there.

### Jobsbolt

I'm going to work on this soon. It's using the batch API but I'll write something which allows to do two things. On-Demand and batch. This would allow me to get things tested immediately and use larger models. Maybe this would be a fun time to rent out a GPU machine. Who knows. This year has been good so far.


I'm starting to think Productsbolt is nearly ready. That's a cool thought

Hi, I'm Alan Reid

Some Blog Posts

Hopefully images are helping

Three weeks going strong

A picture says a thousand words

Variants fixed! Thank goodness for LLMs

A bug when finding new links was found

Productsbolt is nearly ready

Developers Worth Knowing

Alan Reid

Phil Collins

Leo Pillar

Books I'd recommend

Road to React

Eloquent Javascript

Follow my updates per month