WEBVTT 00:00.200 --> 00:01.240 And week three. 00:01.280 --> 00:02.680 Day four is a go. 00:02.720 --> 00:03.880 Let's get started. 00:03.920 --> 00:08.000 So last time we did a stock picker and we just have a little tiny bit more to put into that. 00:08.000 --> 00:10.560 And then we're going to move on with our next project. 00:10.600 --> 00:12.320 A developer agent. 00:12.600 --> 00:14.640 But first another repetition. 00:14.640 --> 00:17.960 I hate to repeat these things, but sometimes it's good to drum it in. 00:18.240 --> 00:20.760 Building a crew project involves five things. 00:20.800 --> 00:23.120 First of all, crew I create crew. 00:23.560 --> 00:28.440 The name of the project to set up your crew, build those directories and those files. 00:28.480 --> 00:33.760 Number two, you find the YAML files for your agents and tasks, and you fill them in to define your 00:33.760 --> 00:35.040 agents and tasks. 00:35.280 --> 00:40.400 Number three, you go to the crew.py module, which is where you actually create the instances. 00:40.400 --> 00:44.960 And you use decorators to identify the agents and tasks that you'll be using. 00:45.000 --> 00:47.320 And then you create your crew itself. 00:47.360 --> 00:53.240 And this is where you can have structured outputs to make sure that the outputs conform to a schema. 00:53.240 --> 00:58.960 And you can use tools, both tools like Serpa, that crew provides for us that run remotely, and then 00:58.960 --> 01:03.060 custom tools that we build locally, like the thing that sends a push notification. 01:03.340 --> 01:10.020 And then number four, you update Main.py to set any inputs so that we can pass something in, configure 01:10.020 --> 01:12.140 the fields that are templated with the curlies. 01:12.500 --> 01:15.100 And finally we run with crew. 01:15.140 --> 01:18.220 I run and off goes our project. 01:18.740 --> 01:22.420 So I now want to cover a feature of crew called memory. 01:22.420 --> 01:28.060 And this is a feature that is a bit more prescriptive, a bit more opinionated in the crew framework. 01:28.100 --> 01:34.140 Memory, of course, is talking about how you provide information, contextual information to Llms each 01:34.140 --> 01:39.180 time you call them, and you can implement that yourself just by storing variables and then passing 01:39.180 --> 01:43.660 them in when you do things like creating tasks so you can do it the sort of manual way. 01:43.820 --> 01:49.180 But the crew framework also comes with some building blocks that lets you use their constructs around 01:49.180 --> 01:52.140 memory out of the box, and that comes with pros and cons. 01:52.180 --> 01:57.020 The pro is that you get up and running quickly, and you can use a lot of the thinking that they put 01:57.020 --> 01:57.820 behind this. 01:58.020 --> 02:03.610 The con is that there's there's a learning curve, and it obscures some of the detail of how prompts 02:03.650 --> 02:05.250 actually work behind the scenes. 02:05.250 --> 02:10.730 So as as with many times when you're adopting a framework like this, it's something to be aware of 02:10.810 --> 02:13.050 the benefits and the trade offs of doing so. 02:13.410 --> 02:18.130 But let's say that we are going to embrace cruise way of handling memory and talk about what it actually 02:18.130 --> 02:18.610 does. 02:18.850 --> 02:24.330 Well, it has five different types of memory, five different frameworks that you can include. 02:24.530 --> 02:26.850 And one of them is called short term memory. 02:26.850 --> 02:34.570 And this is just about storing recent interactions using a vector database in in a in a rag way. 02:34.570 --> 02:37.210 If you're familiar with retrieval augmented generation. 02:37.210 --> 02:40.650 And you don't need to be for this course because we're just going to put the code in there and see it 02:40.650 --> 02:41.130 run. 02:41.130 --> 02:43.250 But if you do no, then this will make more sense. 02:43.250 --> 02:49.850 So this will allow agents to access recent relevant information when they are currently executing. 02:50.450 --> 02:56.330 And then a different concept called long term memory is when more important information is stored in 02:56.330 --> 03:03.010 a SQL database for longer term recall to build up knowledge over over a longer period of time. 03:03.610 --> 03:08.510 And then there's something called entity memory that's very similar to short term memory, actually. 03:08.550 --> 03:15.310 It's it's basically when there's things about people, places and concepts, then those can also be 03:15.310 --> 03:22.430 stored in a Rag database for vector based similarity search and to be included in the context. 03:22.830 --> 03:26.110 And then there's crew describes this as another kind of memory. 03:26.110 --> 03:27.230 But I think it's a bit misleading. 03:27.230 --> 03:33.990 I think what they call contextual memory is just a sort of umbrella term for the short term, long term 03:33.990 --> 03:42.070 and entity memory that can all together be queried and passed in as context when prompting an LM and 03:42.070 --> 03:43.910 crew abstracts all this away from you. 03:43.910 --> 03:48.190 So it's just going to be a few lines of code to have all of these types of memories running. 03:48.390 --> 03:50.430 But in doing so, as I say, big benefit. 03:50.470 --> 03:53.070 A lot of work happens behind the scenes. 03:53.070 --> 03:57.030 And perhaps also there the trade off is that you've got less visibility into it. 03:57.030 --> 04:01.270 So if things don't go the way you want, then it's a bit harder to debug and figure out what's going 04:01.270 --> 04:01.630 on. 04:01.910 --> 04:08.500 And then there is another kind of memory called user memory, which is to store user specific information. 04:08.500 --> 04:14.940 And actually, at least as of now in Crewe, this is a concept that they support and have some frameworks 04:14.940 --> 04:21.300 around, but it's mostly left up to you to be querying user memory and then inserting it into the prompt 04:21.300 --> 04:22.900 or providing it at the right time. 04:22.900 --> 04:28.020 So user memory is a bit of an odd one out here, and I suspect that they're looking to build more into 04:28.020 --> 04:28.940 that in time. 04:29.260 --> 04:33.180 And for now, for the code that we're about to look at, we're just going to really look at contextual 04:33.180 --> 04:33.540 memory. 04:33.540 --> 04:40.300 So short term long term and entity memory and seeing how we can incorporate that into our stock picker 04:40.300 --> 04:40.980 solution. 04:41.420 --> 04:44.540 And so we're back in the stock picker project in cursor. 04:44.540 --> 04:47.260 And we are looking at the crew.py module. 04:47.580 --> 04:53.100 And I'm going to start by putting in some new imports in here, which are interesting ones from crime 04:53.100 --> 04:53.460 memory. 04:53.460 --> 04:58.500 We're going to import long term memory short term memory and entity memory. 04:58.500 --> 05:00.180 The three types we'll be working with. 05:00.220 --> 05:03.660 And I do believe you can also have user memory in there too. 05:03.700 --> 05:06.220 But then you have to manually manage it yourself. 05:06.800 --> 05:13.320 And then from my memory storage rag storage, we're importing a class called Rag storage for vector 05:13.320 --> 05:14.880 based retrieval. 05:15.160 --> 05:24.120 And with that we're also going to import from long term memory SQLite storage, long term memory SQLite 05:24.160 --> 05:26.720 storage object like that a class. 05:26.920 --> 05:27.400 All right. 05:27.400 --> 05:30.640 So that's a few things that we're now going to put to good use. 05:30.640 --> 05:37.080 We're now going to go to the crew function the function that creates the crew within this this module. 05:37.080 --> 05:39.320 And you can see where we made our manager. 05:39.320 --> 05:41.320 And we've got a few more things to make. 05:41.360 --> 05:48.120 As as cursor is trying to prompt us, they're trying to ignore cursors, insisting let's do it ourselves. 05:48.320 --> 05:48.720 All right. 05:48.720 --> 05:55.960 So we are going to to want to be creating a short term memory a long term memory and an entity memory. 05:55.960 --> 05:58.280 And we're going to do them one by one. 05:58.280 --> 06:04.440 So let's start by saying that the short term memory, which is the one that we'll begin with, short 06:04.480 --> 06:08.780 term memory, trying not to press a tab or it's going to fill it all in for me. 06:08.780 --> 06:10.820 But I'm going to I'm going to do it right here. 06:10.860 --> 06:14.940 Short term memory is going to be something which has rag storage. 06:15.340 --> 06:22.460 We we come up with a provider which is OpenAI, and a model, an embedding model to generate vectors 06:22.460 --> 06:23.420 from text. 06:23.420 --> 06:24.620 And we'll be using this one. 06:24.620 --> 06:27.860 And you can substitute in whatever models you would like here. 06:27.860 --> 06:29.220 It's going to be short term. 06:29.220 --> 06:35.380 And we give it a path to where we'd like it to create that memory, that vector store as memory. 06:35.620 --> 06:41.540 And it will use chroma as it happens, something which people who've taken my, uh, alarm engineering 06:41.540 --> 06:42.540 course know very well. 06:42.540 --> 06:43.460 I love chroma. 06:43.700 --> 06:47.180 All right, so that's the short term memory. 06:47.340 --> 06:50.660 Let's also create some long term memory. 06:50.660 --> 06:51.620 So here it is. 06:51.620 --> 06:57.260 The long term memory is going to be just simply creating an instance of this of this class long term 06:57.260 --> 06:58.820 memory SQLite storage. 06:58.820 --> 07:01.020 And we'll give that also a place to go. 07:01.060 --> 07:04.660 We'll make a database file also in the same directory. 07:04.980 --> 07:07.140 So that's the long term memory object. 07:07.500 --> 07:13.850 And now finally, thirdly, we're going to create a, uh, entity memory object. 07:13.850 --> 07:15.810 So entity memory. 07:16.290 --> 07:17.250 And there it is. 07:17.450 --> 07:19.450 So oh I see. 07:19.490 --> 07:19.730 Hang on. 07:19.770 --> 07:20.570 We've got two of them. 07:20.570 --> 07:21.530 Let's do that. 07:21.930 --> 07:22.810 Looks like that's good. 07:22.850 --> 07:25.250 So entity memory is going to be an entity memory object. 07:25.250 --> 07:28.930 It's also going to be a rag storage object. 07:28.930 --> 07:33.770 We give the provider and the embeddings model and we put it in the memory folder. 07:33.970 --> 07:40.370 So here we have our three types our short term memory our long term memory and our entity memory. 07:40.530 --> 07:44.530 And now we get to the place where we create our crew. 07:44.850 --> 07:47.210 And now it's going to be very challenging. 07:47.250 --> 07:48.570 It's not going to be challenging at all. 07:48.610 --> 07:51.170 We're going to say memory equals true. 07:51.890 --> 07:54.810 And we are going to then just do exactly that. 07:54.850 --> 07:57.490 So crew was a little bit off base with the memory equals part. 07:57.490 --> 07:58.690 But it's got the rest of it right. 07:58.690 --> 08:00.330 That's all you need to do. 08:00.490 --> 08:04.970 You set the long term memory the short term memory and the entity memory. 08:04.970 --> 08:07.010 And we are almost done with memory. 08:07.010 --> 08:08.610 Just just as simple as that. 08:09.170 --> 08:14.510 And I said almost because there is just one or rather two very small extra changes we need to make. 08:14.550 --> 08:21.910 We need to go back up in the module to where we created these agents the trading company, the financial 08:21.910 --> 08:23.430 researcher and the stock picker. 08:23.550 --> 08:25.310 And we need to give them memory. 08:25.350 --> 08:29.830 Now what we want is for the trading company finder to have memory. 08:29.830 --> 08:32.190 And we just do it by saying memory equals true. 08:32.510 --> 08:36.390 We don't actually want the researcher to have memory because we want it to go and do research every 08:36.430 --> 08:36.950 time. 08:37.110 --> 08:41.830 But we do want the stock picker to have memory, because we don't want it to recommend the same thing 08:41.870 --> 08:42.550 more than once. 08:42.550 --> 08:48.470 And I don't know if you remember, but in the prompts, in the YAML files, I said a couple of times, 08:48.790 --> 08:52.590 uh, don't, don't recommend the same stock twice and things like that. 08:52.590 --> 08:56.190 And, and surface new companies for the for the trading company finder. 08:56.190 --> 09:01.070 And that would normally be the final change you need to make is go back and make sure that your instructions 09:01.070 --> 09:05.750 and your YAML files are very clearly making sure that it will take advantage of memory. 09:05.750 --> 09:11.580 Because remember, whilst memory, these abstractions are trying to make memory seem quite magical and 09:11.580 --> 09:12.780 taken care of for you. 09:12.780 --> 09:19.660 At the end of the day, memory just means more stuff shoved into the prompt, more relevant context 09:19.700 --> 09:24.220 put into the prompt so that when you call an LLM it has knowledge. 09:24.260 --> 09:30.220 It's in the input is included information about prior conversations or about prior information that 09:30.220 --> 09:31.020 it retrieved. 09:31.340 --> 09:34.660 So with that we have set up set up the memory. 09:34.820 --> 09:39.860 And we are now going to bring up our terminal and then run this. 09:40.220 --> 09:43.900 And so as usual I go into already in the stock picker. 09:43.900 --> 09:47.540 So all I have to do is type a query I run and we'll be good to go. 09:47.740 --> 09:49.060 Let's see what happens. 09:49.260 --> 09:55.420 So just right off the bat, we expect it to be able to take advantage of memory without needing anything 09:55.420 --> 09:55.820 more. 09:55.980 --> 09:59.540 What we should see is it should create a memory directory in here. 09:59.540 --> 10:02.260 And it just has there is a memory directory. 10:02.460 --> 10:04.020 Stuff is going on. 10:04.060 --> 10:10.340 My computer is hard at work and I can see already within memory there is a Croma database that's being 10:10.340 --> 10:10.980 created. 10:10.980 --> 10:12.960 There is a long term memory. 10:12.960 --> 10:18.480 If I expand this, there is indeed a database that's been a SQLite database that's been created there. 10:18.480 --> 10:26.440 And so things are happening and we can see that, that, uh, various companies are being surfaced by 10:26.440 --> 10:33.320 the market watcher and more is going on and we will, uh, let this thing run. 10:33.360 --> 10:37.640 My computer's hard at work, and I will see you when we have a conclusion. 10:38.080 --> 10:39.640 And that definitely took a bit longer. 10:39.640 --> 10:43.040 It was going around around the houses a little bit, but it completed it. 10:43.040 --> 10:44.760 Recommended Microsoft this time. 10:44.880 --> 10:46.680 Remember, don't use this for real decisions. 10:46.680 --> 10:51.080 But it was, uh, entertaining to see it at work and bouncing around between the different agents. 10:51.200 --> 10:56.280 And whilst we don't have as much visibility into what's happening in terms of its use of memory and 10:56.280 --> 11:02.320 what context got provided, we can see that it's certainly built and populated different data stores 11:02.360 --> 11:09.320 and both the short term memory, the long term memory and the entity memory in the Rag chroma data store 11:09.320 --> 11:10.480 that's been created there. 11:10.520 --> 11:15.550 And the main point I want to get across is that, of course you can see the benefits of what this brings 11:15.550 --> 11:15.750 us. 11:15.750 --> 11:23.590 It was so easy to set up quite a complex situation, multiple types of memory with both vector similarity 11:23.590 --> 11:25.710 queries and SQL queries too. 11:25.750 --> 11:27.430 And we didn't need to know anything about it. 11:27.470 --> 11:29.830 We simply created the objects. 11:29.830 --> 11:33.950 The short term long term entity memory objects passed them in. 11:34.070 --> 11:40.190 And then we told our agents we turned memory on by saying memory equals true for the agents that we 11:40.190 --> 11:41.910 wanted to remember things. 11:42.350 --> 11:45.510 And that is a wrap on the stock picker project. 11:45.550 --> 11:49.350 We saw a lot of different aspects of Q AI with this project. 11:49.390 --> 11:52.150 We reminder, we saw structured outputs. 11:52.310 --> 11:55.870 We saw our own homegrown tool as well as Serpa. 11:56.190 --> 12:01.750 We also used the not the sequential but the hierarchical process. 12:01.950 --> 12:09.750 And now we have added in the memory feature all, all the three main types of memory in there as well. 12:09.790 --> 12:14.390 And that is a nice tour of a lot of the functionality in crew.