WEBVTT 00:00.040 --> 00:00.480 Okay. 00:00.520 --> 00:02.920 Next up, a super important integration. 00:02.960 --> 00:05.360 A nice and easy one that you're gonna enjoy. 00:05.560 --> 00:09.160 It's a powerful platform that is called Fire Crawl. 00:09.440 --> 00:16.160 And this is a platform that supports a bunch of different web related activities web scraping, web 00:16.160 --> 00:22.880 crawling, looking for links on a web page, and parsing a web page and running searches. 00:23.240 --> 00:28.760 Turn websites into LM ready data is what it says in the headline, and that sounds like a jolly good 00:28.760 --> 00:29.560 thing for us to do. 00:29.920 --> 00:32.000 And the good news is that it is. 00:32.040 --> 00:33.520 It has a free tier. 00:33.560 --> 00:34.680 It has a free plan. 00:34.840 --> 00:38.880 Uh, that is uh, I'm in the UK, so we're getting we're getting, uh, pounds here. 00:39.040 --> 00:43.920 Uh, wherever you are, you will get the, uh, the right, um, the right currency. 00:44.160 --> 00:46.600 Uh, and it's very, very popular. 00:46.640 --> 00:48.480 A great thing for, uh, for us to do. 00:48.480 --> 00:52.040 And so we are going to sign up for a fire crawl plan. 00:52.080 --> 00:55.360 Fire crawl is the the website here. 00:55.360 --> 01:02.070 And so sign up, uh, and, uh, I will continue with Google and I will authenticate with my Google 01:02.070 --> 01:04.910 account and in we come. 01:04.950 --> 01:07.670 And I'm going to say they've done their OAuth two. 01:08.350 --> 01:09.270 As you now know. 01:09.310 --> 01:13.750 You know the pain that Fire Crawl had to go through to be able to put up that box and the redirect URL 01:13.790 --> 01:14.390 they will have got. 01:14.390 --> 01:14.870 Right. 01:14.990 --> 01:16.630 Um, and here we are. 01:16.670 --> 01:22.990 We've come in uh, and uh, email has been verified and, uh, these are things to get started that 01:22.990 --> 01:25.630 we won't do any of this, uh, we will get ourselves set up. 01:25.630 --> 01:26.310 Continue. 01:26.550 --> 01:30.750 Um, and we're about to complete the setup of our fire crawl account. 01:30.790 --> 01:34.230 Next up, you have a terms of Service screen to to review. 01:34.470 --> 01:39.430 Uh, for example, the description is that fire crawls an API that converts any website into LLM friendly 01:39.430 --> 01:45.190 data, provides tools to extract structured data from web pages, ensuring data is clean and ready for 01:45.190 --> 01:46.470 use in AI apps. 01:46.710 --> 01:49.910 And you'll have to scroll all the way through all of this, and then you'll be able to turn that to 01:49.950 --> 01:51.510 yes, and then you'll be able to continue. 01:51.550 --> 01:57.030 We'll show you a little bit of code that you can ignore and proceed on to, to the, uh, the dashboard 01:57.030 --> 01:57.390 screen. 01:57.390 --> 02:01.170 It's actually it starts by saying invite your team, which we will not do. 02:01.530 --> 02:04.050 Personal is fine and choose your plan. 02:04.050 --> 02:11.490 We will get started with the free one allows us to scrape 500 pages, get started and it says welcome 02:11.490 --> 02:12.410 to five crore at the top. 02:12.410 --> 02:13.250 And here we are. 02:13.290 --> 02:19.050 Here we are at the dashboard screen and you can see that we have, uh, all, all sorts of things, 02:19.090 --> 02:25.690 the different things that it can do is scrape, which is to get data from a website search, which is 02:25.690 --> 02:30.890 given a search term, be able to, to look for, for do a, do a, basically like a Google search crawl 02:30.890 --> 02:35.490 is given a web page, find all of the links out different pages and get that data. 02:35.490 --> 02:36.930 And then the agent. 02:36.930 --> 02:38.770 Everyone has agent functionality these days. 02:38.770 --> 02:41.610 We're going to be building that kind of thing using these. 02:41.850 --> 02:44.970 There is an MCP integration that we'll learn all about tomorrow. 02:45.130 --> 02:49.490 Uh, but uh, for now, what we really need is the API key. 02:49.810 --> 02:56.600 That is the thing at this point to be copying into your clipboard, uh, so that we can use it in N810. 02:56.640 --> 02:58.480 Okay, so now we go back to N810. 02:58.520 --> 03:02.480 We go back into my first folder and let's go back to the Deep Seek project. 03:02.520 --> 03:03.440 Here it is. 03:03.720 --> 03:04.000 Okay. 03:04.080 --> 03:05.360 First thing a little bit confusing. 03:05.360 --> 03:11.720 I just want you to create a new node and type fire crawl and look at this fire crawl node. 03:11.720 --> 03:16.080 What you'll see here, probably if you're like me, is that there's this install node button. 03:16.240 --> 03:21.040 And this is, is going to install this as, as a node that we can use in cloud. 03:21.080 --> 03:24.840 And I'm going to press this so that we install the fire crawl node. 03:25.080 --> 03:26.400 And that is done. 03:26.640 --> 03:29.520 And that package has been installed. 03:29.880 --> 03:31.840 Um okay great. 03:32.080 --> 03:34.200 Now let's use it okay. 03:34.240 --> 03:39.760 And now I'm thrilled to introduce to you the final piece of a new learning for today, a really big 03:39.760 --> 03:42.840 one, uh, topic called structured outputs. 03:43.040 --> 03:47.560 So we're about to add here, fire crawl to to run a web search. 03:47.720 --> 03:51.440 And it turns out that fire crawl isn't available in as a tool. 03:51.440 --> 03:55.830 There's no fire crawl tool, so we can't just simply plug it into the AI agent. 03:55.830 --> 03:58.670 We have to put it in as a node to come after the output. 03:58.670 --> 04:03.870 So the agent will make some output and that will become the input into driving our web search. 04:04.070 --> 04:10.230 But there's a problem here, which is that LMS just generate output content in an unstructured way. 04:10.230 --> 04:15.190 It doesn't conform to any particular JSON spec, it's just the response that the LLM has. 04:15.390 --> 04:16.670 And that's not what we want here. 04:16.670 --> 04:21.030 We want it to produce some JSON according to a particular format. 04:21.230 --> 04:24.990 And luckily there's a way to do it and it's absolutely crucial. 04:24.990 --> 04:29.270 And I, I use it all the time, as do do all practitioners in the field. 04:29.310 --> 04:33.750 It's this technique called structured outputs and it's built into n810. 04:33.790 --> 04:35.790 And that's what we're going to use right now. 04:35.790 --> 04:40.150 So for any AI agent like this you can double click on it, bring it up. 04:40.150 --> 04:44.030 And there is this require specific output format. 04:44.030 --> 04:45.390 We're going to turn that on. 04:45.790 --> 04:47.230 And that's so crucial. 04:47.230 --> 04:54.540 And once we've done that look at this A new a new thing has appeared in in this this this node cluster. 04:54.700 --> 04:59.060 Uh, and it is something to connect an output parser to. 04:59.260 --> 05:03.100 And one of the output parses is called a structured output parser. 05:03.100 --> 05:04.940 And that's the one we want right now. 05:05.140 --> 05:12.500 And it's one that says hey LM whatever you produce, it's got to conform to a particular type of JSON. 05:12.500 --> 05:14.060 We need you to do that. 05:14.060 --> 05:15.900 And as a result, it will. 05:15.980 --> 05:21.460 And it's really cool that Nan allows us to do it by just giving some example JSON. 05:21.620 --> 05:27.260 And then it will guarantee that the LM will generate an output that conforms to this. 05:27.260 --> 05:28.900 So look, I'm going to have a bunch of JSON. 05:28.900 --> 05:32.140 I'm not going to have what they put in as the default state and cities. 05:32.260 --> 05:44.460 Instead I'm going to have uh search query, colon, uh, internet search or whatever. 05:44.860 --> 05:47.740 Uh, it's an internet search perhaps would be a bit better. 05:47.900 --> 05:49.100 Uh, doesn't matter. 05:49.100 --> 05:50.300 But this is an example. 05:50.300 --> 05:56.680 Very simple piece of Jason Jason with only one key value pair search, query and internet search. 05:56.680 --> 06:02.440 And I'm gonna I'm gonna make this this this is going to be the, the output, uh, query structure. 06:02.440 --> 06:07.080 And that's saying to the LM, whatever you generate, you can generate whatever you want as long as 06:07.080 --> 06:08.720 it conforms to this. 06:08.720 --> 06:12.280 Jason, this type of Jason generates it from this Jason example. 06:12.560 --> 06:13.080 Okay. 06:13.120 --> 06:15.720 And now we come back here and you can see that that's there. 06:15.720 --> 06:18.680 And that is what this agent is going to generate. 06:18.680 --> 06:21.440 And just to keep it coherent I'm going to double click here. 06:21.440 --> 06:23.720 And I'm going to add in a system prompt. 06:23.760 --> 06:26.240 Come here to Options System message. 06:26.240 --> 06:30.920 And we're going to change your a helpful assistant to be something that makes its task slightly more 06:30.920 --> 06:31.480 clear. 06:31.520 --> 06:32.000 Here we go. 06:32.040 --> 06:40.720 Let's say something like you come up with search queries for an internet search. 06:43.400 --> 06:48.920 To best, uh, answer the user's question. 06:49.600 --> 06:50.630 That's what we'll say. 06:50.630 --> 06:53.870 You come up with search queries for an internet search, or we'll say search query. 06:54.110 --> 07:01.190 With a search query, you come up with a search query for an internet search to best answer the user's 07:01.190 --> 07:02.110 question. 07:02.390 --> 07:02.990 All right. 07:02.990 --> 07:04.470 That's how we're setting it. 07:04.510 --> 07:05.470 Save that. 07:05.510 --> 07:06.510 Come back here. 07:07.350 --> 07:09.350 Uh, we didn't need to save it, but but I did. 07:09.510 --> 07:13.390 Uh, and with that, we're now going to try this out before we add the parser on. 07:13.390 --> 07:14.750 Let's just see what happens. 07:14.790 --> 07:15.350 Okay. 07:15.390 --> 07:18.230 So I'm going to come in here, I'm going to I'm going to open the chat. 07:18.230 --> 07:31.030 And I'm going to say, uh, I'd like to find out what are the best LMS to be using in 2026. 07:31.230 --> 07:35.750 And that is the message that we are going to give to our model. 07:36.070 --> 07:40.270 And it's got that system prompt that tells it that it's going to try and make a search query and it's 07:40.270 --> 07:40.750 run. 07:40.750 --> 07:43.670 And if we double click on this what you'll see has come out. 07:43.710 --> 07:45.150 We'll switch to the JSON view. 07:45.310 --> 07:52.020 You will see that it's produced in under output something which conforms to the structured outputs that 07:52.020 --> 07:52.980 we requested. 07:52.980 --> 07:56.660 It conforms exactly, and it always will. 07:56.860 --> 07:58.540 And I can't emphasize enough. 07:58.580 --> 08:03.420 This is an incredibly powerful trick to use with LMS. 08:03.660 --> 08:10.580 LMS give us this this rich ability to make a nuanced decisions based on their inputs, which is great. 08:10.580 --> 08:15.980 But one of the downsides is that what they produce could be they can they can produce tokens that are 08:15.980 --> 08:17.980 very unstructured and can be anything. 08:17.980 --> 08:22.540 So trying to turn that into an organized workflow can be quite challenging, particularly if you're 08:22.540 --> 08:23.580 not using tools. 08:23.780 --> 08:25.380 This is the answer. 08:25.620 --> 08:27.580 Always use structured outputs. 08:27.580 --> 08:28.100 Turn on. 08:28.100 --> 08:29.700 Require specific format. 08:29.740 --> 08:31.620 Fix a JSON format. 08:31.620 --> 08:37.380 Make sure your system prompt is consistent so it's clear what the LM has to do, and your outputs will 08:37.380 --> 08:44.140 conform to a structure like this with a nice search query in this case, but with whatever you want 08:44.140 --> 08:44.860 it to do. 08:44.900 --> 08:49.770 And this is now ideal for us to hook up to the next step in the process. 08:49.770 --> 08:53.130 And of course, the next step in the process is going to be fire crawl. 08:53.250 --> 08:55.890 So we add in Fire Crawl right here. 08:56.050 --> 09:01.650 We are going to say search and optionally scrape web results. 09:01.650 --> 09:02.450 Here it is. 09:02.770 --> 09:06.130 Credentials connect with is fire crawl account. 09:06.130 --> 09:10.450 So you've pasted in your your API key as I have. 09:10.490 --> 09:14.330 And you've got the connection tested successfully in here. 09:14.490 --> 09:16.210 I hope all is good. 09:16.370 --> 09:19.090 Uh, so we're going to, uh sorry. 09:19.130 --> 09:20.050 Come on in here. 09:20.450 --> 09:25.530 So the field that needs to be specified is what is it going to query on. 09:25.650 --> 09:26.770 And of course it's right here. 09:26.770 --> 09:31.010 It's the search query because this is exactly what came from the AI agent. 09:31.050 --> 09:35.130 You can either type in the expression yourself or you can drop it in there. 09:35.170 --> 09:36.930 It switches to expressions automatically. 09:36.930 --> 09:41.010 It is of course JSON output dot search query. 09:41.010 --> 09:42.690 And there is the example. 09:42.890 --> 09:45.530 And everything here looks good. 09:45.990 --> 09:49.310 Uh, okay with that, let's come back here. 09:49.470 --> 09:52.070 Let me copy this message that I had before. 09:52.110 --> 09:55.190 Let's start a fresh chat using this. 09:55.230 --> 09:56.350 Paste this in. 09:56.510 --> 10:00.510 And now everything runs off it goes. 10:02.310 --> 10:05.270 It's gone to deep seek again through open router. 10:05.310 --> 10:08.150 The structured output came out and it ran. 10:08.150 --> 10:10.390 Let's double click on here to see what came out. 10:10.430 --> 10:18.670 What came out is a bunch of URLs, titles and descriptions of of information that it retrieved based 10:18.670 --> 10:19.710 on that. 10:19.710 --> 10:23.150 That question about finding 2026 models. 10:23.150 --> 10:28.030 And you can see it's got some good links there to a Forbes article, to Zapier, to to all sorts of 10:28.030 --> 10:28.430 things. 10:28.430 --> 10:30.310 Here it is working. 10:30.470 --> 10:33.750 We have just scraped the internet using Fire Crawl. 10:33.870 --> 10:40.030 Um, based on a general question that we asked, and we used an AI agent to turn that question into 10:40.030 --> 10:42.550 a search term for fire crawl. 10:42.550 --> 10:45.070 That is a great integration together.