Local AI Web Scraping with n8n & Oxylabs – Workflow Breakdown!

Added: 2026-05-10

[00:00:00]strong match and everything we see right here is actually generated from our local AI. Today we're going to be taking a look at how web scraping tools can work in conjunction with workflow automation like NADN and local AI to power workflows that can really accommodate a very wide variety of use cases. For today's workflow in specific, we're going to be looking at a tool that I built personally through NADN Olama and using Oxyabs web scraping capabilities, who are also the sponsor of today's video. So, thank you to Oxyabs for sponsoring this video and allowing me to build something cool that I'm really honestly excited to showcase.

[00:00:37]This is an idea I've had for a while.

[00:00:39]And the whole point of today's video is to showcase how this works to create what we see right here. So, this is a drop shipping demo. Now, oftentimes when you hear the word drop shipping in a YouTube video, you're probably having red flags going off thinking, "Oh no, this person is about to try to sell me a course." But I'm very happy to report that for today's video, our demo is going to be completely opposite to that.

[00:01:01]Because we're going to build something or showcase something that shows if an Amazon product we're looking at is drop shipped, and if so, just give us the actual comparable link to find that listing directly from the source to see how much additional price is being added on, etc. So, we're going to be using this for good. And again, thank you to Oxyabs for sponsoring this video. And in conjunction with that, because influencer status, if you go to oxyabs.io/bjan in all caps, and probably not in all caps as well, you will see that you can get 20% off and you can actually use this promo code right here to get a free trial without needing to enter credit card data. So, thank you very much to Oxyabs for the influencer status sponsor. And now, let's get into it and actually talk about today's video. We'll do some simpler demonstrations and then we'll look at this big very dense N8 workflow and we'll talk about the local AI component as well. Now, in order to properly really introduce today's system, I think it's best to look at each of the individual components that are going to make up a specific example that we're showcasing here, but beyond that, how we can actually use something like Oxyabs to fit into a workflow with NADN and a local AI model, which in this case is Quen 3.635B- 6 35B- A3B and we'll talk a little more about that later on and why specifically that model was chosen. So in this case, OxyLabs is essentially acting as the web layer where it's finding structured web data for us. Well, it's finding unstructured web data, I should say, and it is structuring it and organizing it for us.

[00:02:30]And it is allowing us to essentially get lots of data in a way that is not going to balloon the context window that we need for our local AI model. On a system like this, which is probably kind of tapped out in running this Quen 3.6 A35B model, there's not a lot of room for additional space in terms of making the context window larger when running this model. So, if we were having this model work with tools where it had to actually go out autonomously on the web itself, we could perhaps encounter issues where certain sites would block us. though for today's video and what Oxyabs is doing is only for public available data. So I should mention that first and foremost but essentially our model would get choked down in context if we had to just rely on it solely. So Oxyabs is giving us that web layer where it's going to find data for us and provide it in a structured manner for our model so that our model can efficiently process the things that Oxyabs handles in lie of having to do all of that work itself which on a resource constrained system like what we're using today would just kind of make this infeasible. Now, additionally to that, and because I don't want to specifically jump into trying to explain this somewhat intimidating looking NAND workflow here, the thing to do is we're going to go into our Oxyab dashboard and actually take a peek at the specific API playground. So, if we go in here, we see that currently my products are the web scraper API, which is what is going to be powering today's video. And if we click on that, we see I've done a bunch of experimentation with this. Of course, though, we also have an explore API playground. And for the purpose of a simple call it smoke test demonstration, this will really allow it to make a lot more sense in what specifically Oxyabs is doing for this video. So, let's showcase a really simple web scraper result that Oxyabs will be running themselves just by us enabling this by giving it the search string we want, specifically picking a website. So, in this case, we could click Google, we could click Amazon. For right now, I think Google will probably be the best service for what we specifically want to showcase. And let me just enter in a search query here. So, for example, let's just look up Bjan Bowen since I guess that would be somewhat pertinent to the video. Now, we can just submit the request right here, which will then go and search Google and showcase us, and we'll see it in the output preview as well. But something additionally we see here is there are a bunch of additional options which actually will modify the request code here which gets shown in a variety of different languages which makes it simple to integrate this into a bunch of workflows not specifically NN. So if you were building something that was more custom in Python you would be able to just copy paste a Python snippet here but you'd be able to see how it looks in this playground initially which just does make things easier. Now, additionally to that, we can also set our specific localization. Certain times, there will be different prices actually shown to different local. And being able to look at a bunch of different ones to culminate in like a specific average price is a very, very important and valuable use case for a lot of industry.

[00:05:29]Now, for today's video, we're not necessarily as focused on something like that, but the ability to set geoloc in and of itself is pretty important. So, for this, let's just type in United States as that where I am. And we see that gets added on to our sample code snippet. Additionally to that, we could select a user agent, which is essentially the kind of browser or system that is being used to search for this result. And there are a whole bunch of other additional options that we can select here. For the simplicity of an initial smoke test demo, I find it's probably sufficient to just have this specific request right here. And as we submit this request, we're going to see that once it shows us our output preview, we're going to actually just see the Google page results for Bjamboan. So if I were to search this exactly right now on Google, this is the exact same thing we would get in our results. Basically a mirror image. The only real difference being that this isn't in dark mode. But I suppose that's okay because this is just a playground result. Additionally though, and something important to see is we get our output right here. So this is more structured output that we're getting. additionally to just the visual of what we see in this output preview and having this is something that really unlocks our ability to use a smaller model with a smaller context window in terms of the local AI powering our experience. So in lie of giving our local AI like a script that will have to go out and try to figure out how to do this which will make it slower because it will have to be using lots of tool calls. The context length will get filled up very quickly when it tries to ingest and parse all of this. Oxyabs basically just gives us the structured code in an output for a specific search query and then we can just pipe that directly into our local model pipeline in something like NADN and it allows it to function more efficiently and I don't want to sound like I'm beating a dead horse but with less of a context window needed and less of a massive beefy system to actually perform functional things. So that is a simple kind of overview TLDDR of what specifically is going on here with this Oxyabs web layer and what value it provides for an example like this. So the playground here shows the request code and then a preview of the parsed output. But in order to tie this into Nadn so that our full-on dropshipip detector demo actually makes a little more sense. I think it'll be cool to just run a very tiny smoke test demo that will almost emulate what's happening here, but then tie it in with the local Quen model that we are running currently on the system.

[00:07:55]So from in N8N right here, we can see this is a very simple and palatable three node workflow, which starts with a manual trigger, which means that in order to actually initiate this, we just click this execute workflow button right here. And now I'm going to do this. And really the slowest part of this whole thing is going to be having the model that is running locally on this laptop actually return a result with what we want it to say. Now from what we can see right here, it's not going to take too long because this is a pretty simple smoke test demo. But let's actually take a look at these nodes and see what specifically just occurred right here and why this ties into actually having a web scraping result that makes things easier on our local model. So if we click on this Oxyabs Google search node, we're going to see right here that we have it essentially running the same thing that we did in the playground dashboard right here where it searched Google for Bjambowen and it found some structured output. It structured it for us and then returned it in this manner.

[00:08:53]So when we go back to the node in nadn right here we can see that in this JSON node we do have structured output here with some specific links that it found pertaining to Bjan Bowen on Google and some additional like images and things like that. Now for the purpose of simplicity and keeping this as a light smoke test demo we're not using a lot of the data here. So we're not using the images and things like that. Think of this more as utilizing the summary that Oxyabs has just produced for us through the web layer right here. So once the OxyLab search result has been completed, we're going to look at the node that follows it, which is being fed that structured information from our web layer, which we can see right here as this node's input. Now, this node, if we notice, does specifically say Olama summarized search results. So if we click back on this, we'll see right here that I am using NAND through Docker. So that's why this has host.doccker internal. But this is just the regular server port for Olama, which runs on 11434. So this is basically just sending this structured data here to our Quen 3.635B model. Additionally, it is also giving it a specific prompt or task to perform with this structured data which if I can scroll in this tiny little window and find it, we will see that it is summarizing Google search results for a developer audience. Return JSON only with the following schema. And then we see that right here. So the output here for this node is actually produced by our local Quen model. So, we got the Oxyabs web search results as input. We told it specifically, okay, you need to summarize this for a developer audience and then e export it or show it as this specific JSON schema, which it did do.

[00:10:31]And all of the results right here in this right-hand pane are actually produced by our local AI model where it used this structured data as well as our instruction to basically just give a little summary of what was found through Google. So this is a really simple smoke test to try to I suppose architecturally understand how all of these things are working together. These things being NAND for the actual workflow orchestration, Oxyabs as the web layer to provide structured web data for us and then our local AI model to use that structured web data as well as a specific task that we denote it should do to produce a result without needing to handle all of that itself which quite frankly would not be able to work with this specific model on this system because the context length that would be necessitated by having to do manual searching and some of the fragile scripts that would have to be written and would be very difficult and significantly more time and work than the really like cheap to use API calls that are being done through OxyLabs. So, with that simple smoke test demo out of the way, let's actually focus on the main event, if you will, which is our Amazon listing drop shipper detection.

[00:11:39]And again, before we look at our drop ship detecting demo, if you do want to try this yourself, you can go to oxyabs.io/bjan io/bjian in all caps and you will get a discount code as well as some free utilization of this without needing to enter a credit card which is always optimal. So now let's do a real life run with our drop ship detector through this rather intricate N8N workflow and I want to showcase this functioning beforehand and then following this we'll actually take a closer look at some of the aspects of this workflow and how it is all working together. So to begin, we're going to actually enable this workflow being that it will be triggered by this web hook which does get triggered by our drop ship detector demo. And this is just a front end running on the local system that is hooked into the NADN pipeline here, if you will. So to begin, we're going to select execute workflow there.

[00:12:27]And this is now ready and listening for us. Now the drop ship detector demo has a sample item that it loads in with. I don't 100% recall what that specific item ID is. I do believe it's a gaming mouse or something, but I have located this which is very popular as it has sold over 6,000 in the past month and it's a laptop cooling stand and it's fairly identifiable just based off of the fact that we can visually see there are three fans there and things of the sort. So, our next step is going to be to copy this specific product ID, which is just shown after that DP text right there. and we're going to paste it in in place of the sample ID, making sure that we don't include extra things which will mess stuff up. And then basically all we're going to do is select run live demo. And as we see right here, the steps are being run through and everything happens fairly quickly until we hit the local model verdict. And the reason for that is basically as follows.

[00:13:21]So we're using a Quen 3.6 35B model on this laptop, which is a very large model for a computer like this. So, it is going to be pretty slow in terms of its response speed. Additionally to that, I have opted to use a vision model here because additionally to just being able to try to match this product to something from AliExpress by just using the item titles, it's also going to go a step further and actually pull the images from the Amazon listing one and two from the AliExpress listing. and then the vision model will be able to compare them and give us feedback on whether or not it actually sees these as the same exact item or not. So that's just an additional level and showcase of what we can actually do with this local vision model in this entire pipeline. So as this will take a little bit of time, we're going to wait for that. But if we do go over to our drop ship detector naden workflow right here, we can see that it is currently working on some vision capability things, but everything else happened pretty quickly. And following that, we'll see our Olama nodes start to go. And again, we will look at these a little more in depth, but I did just want to give some form of actual like live demo prior to looking into the workflow. So, as we see right here, we did just receive our result with a verdict of a likely drop ship.

[00:14:36]So, to begin, I suppose we should just read through that. We have a confidence score as well. Strong visual evidence confirms the Amazon listing is the exact same physical product as the AliExpress candidate. The fan layout, mesh pattern, and branding placement are identical.

[00:14:51]Although the AliExpress title is generic, the visual match is definitive proof of the same source. And this is specifically why I wanted to showcase this with a local vision model because it is actually able to compare the images from the product listings, which really adds a level of I suppose just impressiveness, I guess could be said for the demo. So, first and foremost, we have some specific evidence right here.

[00:15:13]Vision analysis confirms a strong match.

[00:15:16]Amazon listing uses this brand which is visible on the AliExpress product page indicating a relist of the same OEM product and there is a pricing differential. So if we scroll down here, this is exactly what we want to see and this is why I wanted to use this with a vision enabled model because we actually get an image analysis as well. So strong match and everything we see right here is actually generated from our local AI and this is the whole point of this and this N8 workflow is essentially Oxyabs gives us the web data in a structured curated way that would not really be possible to get with this local model. It just would the context would tap out the tools would be unreliable and it would be really a nightmare to try to consistently get that over and over again. So that serves us up the data. Then NAN just kind of pipes everything together as the orchestration layer and then our local model performs all of the reasoning and assessment for us just with the data that it's given from the structured way that we're getting it from the Oxyab scraping. So let's see right here we have vision confidence 95% model is the newest Quen 3.6 which is what we're using the 35B A3B which is just that model label when using Olama which is what we are. We have an identical fan configuration. Honeycomb mesh pattern on the surface is identical. The logo is located in the exact same position in the top right corner of the device. And the side vent cutouts and overall chassis shape match perfectly. Now, this is a relatively potent vision model. The Quen 3.5 3.6 family of models has pretty robust vision capabilities, especially considering the size. So finally, we're going to click on the link we've received right here, which is our likely source, and we get it exactly. So this is actually it almost looks like it just used the the Amazon image twice. But I mean, the point is that it is specifically this item. And we also have a specific source price and an Amazon markup right here, which really is kind of cool. So, this is something that at the beginning of this video, I believe I said like if you hear like drop shipping in a YouTube video, you're going to roll your eyes and think like, "Oh no, I'm about to be sold like a course." But this is almost like the exact opposite of that and something that I'm very like I just think this is cool because it uses local AI and Oxyabs web scraping and NAD to actually build something that's functional and a cool demonstration use case because this is just one of a ton of different demonstrations that could actually be built with having curated web data like this. So I mean let's just I suppose we could put these side by side and okay so 2518477 and source price $1791. So it's interesting and you have to remember because we are using Oxyabs as the mobile web scraping or not mobile as the web scraping the price that it sees may be different depending on where it is actually accessing that site from versus me being in my local. So that can be attributed to that. But regardless, we do see an Amazon markup, the Amazon price, etc. So, now let's just take a look at this big N8 workflow, which is not as intimidating as it may seem as a lot of this is just wiring different things together. The core logic here is pretty simple where we fetch information, we take our product, we have Oxyabs find sites and return structured data pertaining to search queries that are actually partially built by our local AI model. We then build the search query to send it to AliExpress and then Oxyab searches AliExpress with that search query that is actually built by our local model which we can see right here is actually referring. So we prepare I suppose I'll give like a simple top level overview but it's not going to be node by node or anything. So first we have our web hook which is just triggered by run live demo here in the web UI. Following that, we take our Amazon product ID and then Oxyab scrapes and searches the page for that specific product ID, returning us structured information about it.

[00:19:18]Following that, we actually ask our local model if this seems like a known brand or it seems like it has indicators of a potential drop ship. Things like a lot of generic terminology in the title instead of a specific well-known brand and then like a identifiable product name. Following that, we have an Olama search query. So this is actually using our local model with the preceding information to actually create a search term for AliExpress that the Oxyabs web scraper in this node will actually search AliExpress using kind of in a similar functional way to how we saw in the web dashboard in the playground when we searched Bjambowen on Google. So following that and it's cool because our local model is actually coming up with the search terminology to search AliExpress, give it to Oxyabs to actually perform that search and it's just coming up with those words based off of seeing the initial Amazon listing that was given to it in that more structured way. So this is kind of how the web layer plays into this and then NN is obviously orchestrating everything. So following that and to simplify, we have a bunch of logic here that tries to pick like the potential AliExpress candidates or items on AliExpress that are more likely to match up with our Amazon listing, assuming that there are signs of this being a drop shipped item. Then following that, we have a bunch of things related to image parsing logic. And this is something that is definitely more fragile in a workflow like this because especially with AliExpress, it's actually kind of difficult to pull images directly from that page because there's often times a ton of unrelated things there. Additionally to that, it's a pretty hostile site in terms of like actually wanting to get structured data from it. So there's always popups coming up for like new user bonus and things of that sort. So it's really kind of messy and it's impressive that we actually can get that data just from the Oxyab scraper. Here we also have some fallback logic in case things are not showing up properly like images. We also have the logic of course to fetch the Amazon image so that our vision model can actually compare them. We have more data preparation and things like that to just give the model the proper payload as we see here. Prepare vision payload. And a lot of this is just kind of a little messy down here. But then we have the vision analysis task which we saw in our detector demo where we actually if we scroll down have this image analysis segment as well. And then really the culmination is just the final verdict which gets sent to our model and it takes a little while. So this is the node that precedes the final of basically everything getting generated and shown in our web UI right here. So the whole purpose of this is just to showcase that this works properly using the three separate items nad for workflow orchestration which is essentially what we see right here. It is wiring everything together and allowing the data to flow through this.

[00:22:08]Following that we have Oxyabs which provides structured parsible web data fed into our model in lie of needing the model to have some delicate script that will try to basically look at the entire page by itself which would choke the context length and especially on a system like this it wouldn't even really be feasible. Then of course we have our local model which is actually performing some impressive tasks here in and of itself where if we go back to some of these earlier nodes we see that it's actually generating the search query to send to Oxyabs to search on AliExpress which we then take that return data and use it to try to pick the candidates that are more likely associated with the product in question that is drop shipped. And it's just all kind of cool how it works together in conjunction.

[00:22:54]And it works together pretty nicely, I suppose I would say. So that is a kind of not TLDDR overview, but a top level overview of how exactly this whole drop ship detector is functioning and how all of the pieces work in conjunction to actually create something that works with a local model as well. So that's going to conclude today's video showcasing basically building something really cool with OxyLabs N8N and a local model in this case Quen 3.635B- A3B through Ola. So if you want to go to oxyabs.io/bhan io/bjan in all caps, you will see this page where you can get 20% off or you can use this code right here and without even entering a credit card, you can get 2,000 scraped results for free just to be able to play with if you want to actually build a workflow like this. They do have a ton of good documentation for building a whole bunch of things. They talk about AI Studio even where you can actually like ask it to create a scraping query for you. So that's kind of interesting because it would definitely simplify some more exotic use cases. Beyond that, there's just some additional like developer docs and things of the sort. So with that, again, thanks to OxyLabs for sponsoring this video and allowing me to showcase some of my NAD skills. Additionally to that, if you have any questions, please feel free to leave them in the comments.

[00:24:16]There will be a link for this in the description as well as a pinned link.

[00:24:19]And thanks for watching.