WEBVTT 00:00.210 --> 00:01.380 -: Hey, welcome back. 00:01.380 --> 00:02.610 And in this short video, 00:02.610 --> 00:04.740 what we're gonna cover and start introducing 00:04.740 --> 00:08.040 is this concept called agents inside a LangChain. 00:08.040 --> 00:10.020 Now, agents are really fascinating, 00:10.020 --> 00:14.640 and they're very good at doing task planning, for example, 00:14.640 --> 00:16.710 and executing lots of different things 00:16.710 --> 00:18.660 in a very dynamic way. 00:18.660 --> 00:20.430 But before we go into any code, 00:20.430 --> 00:21.810 I think it's worthwhile us thinking 00:21.810 --> 00:25.020 about some high-level concepts that LangChain provides. 00:25.020 --> 00:27.780 So you've got this idea of an agent action, 00:27.780 --> 00:31.110 which is what you might decide a tool that's gonna be used, 00:31.110 --> 00:32.940 and we'll talk about tools in a bit 00:32.940 --> 00:34.350 when the agent is finished. 00:34.350 --> 00:36.300 So there's this idea of the eventually, 00:36.300 --> 00:38.130 the agent is gonna finish, 00:38.130 --> 00:40.410 and that it will return some result. 00:40.410 --> 00:43.110 Now there's some series of intermediate steps. 00:43.110 --> 00:45.360 And if I scroll down a bit and show you, 00:45.360 --> 00:47.610 you can almost think of an agent like this, 00:47.610 --> 00:50.220 where the agent gets an action, 00:50.220 --> 00:52.920 and when the action isn't a finish, 00:52.920 --> 00:55.950 we then run the next action. 00:55.950 --> 01:00.360 And then we get an observation, and with the observation, 01:00.360 --> 01:03.510 we then decide to get another action. 01:03.510 --> 01:08.400 So we've provided it almost like an endless loop of actions, 01:08.400 --> 01:11.400 observations, getting the next action, 01:11.400 --> 01:13.680 and then whilst the next action isn't a finish, 01:13.680 --> 01:15.720 we keep running in this endless loop. 01:15.720 --> 01:17.370 So that's a little bit on agents 01:17.370 --> 01:19.470 and the concepts that are happening here. 01:19.470 --> 01:21.330 Now, as you're probably all aware of, 01:21.330 --> 01:23.820 agents use something called tools. 01:23.820 --> 01:25.650 And tools, if you can think of it, 01:25.650 --> 01:27.630 are basically just Python functions, 01:27.630 --> 01:30.060 or some way to interact with the environment. 01:30.060 --> 01:31.740 Now if we have a look at tools, 01:31.740 --> 01:33.660 there's a variety of different ways 01:33.660 --> 01:35.340 that you can define tools. 01:35.340 --> 01:37.650 You've got some default tools that are provided to you 01:37.650 --> 01:40.140 from the LangChain documentation, 01:40.140 --> 01:41.910 but essentially, in essence, 01:41.910 --> 01:44.280 a tool is basically a function 01:44.280 --> 01:47.820 that you can have an input into and there's an output 01:47.820 --> 01:51.780 that gets added, right, from using the tool. 01:51.780 --> 01:54.270 A LangChain provides some really great ways for us 01:54.270 --> 01:55.950 to create our own tools. 01:55.950 --> 01:57.630 And so what we're gonna do is have a look at 01:57.630 --> 01:59.400 how we can create tools, 01:59.400 --> 02:02.880 and how we can attach those tools to LangChain agents. 02:02.880 --> 02:05.130 And then the agent can then use those tools 02:05.130 --> 02:06.420 to answer questions. 02:06.420 --> 02:08.790 So let's have a look at this notebook. 02:08.790 --> 02:10.230 And to start with, 02:10.230 --> 02:12.930 we're gonna use the LangChain_community.tools, 02:12.930 --> 02:14.550 and you've got two tools here, 02:14.550 --> 02:16.530 so I'm just gonna make sure my notebook's running 02:16.530 --> 02:17.970 and I'll just run that. 02:17.970 --> 02:19.800 So let's have a look at this. 02:19.800 --> 02:23.280 So what we're gonna do is have a look at this one here. 02:23.280 --> 02:26.610 All right, so in this example, we're using a community tool 02:26.610 --> 02:29.790 which allows us to query Wikipedia. 02:29.790 --> 02:32.070 And you can see there's this API wrapper. 02:32.070 --> 02:34.050 And then we have this tool 02:34.050 --> 02:36.780 that's being made from the Wikipedia query run. 02:36.780 --> 02:39.900 And this eventually will basically make a tool. 02:39.900 --> 02:42.000 Now we know it's a tool in LangChain 02:42.000 --> 02:43.890 because if it's got a name, a description, 02:43.890 --> 02:46.860 and some input arguments, you can see, 02:46.860 --> 02:48.840 the tool name is called Wikipedia. 02:48.840 --> 02:51.660 This is the description, you've got the query. 02:51.660 --> 02:54.150 So this is the args that will go in 02:54.150 --> 02:57.630 and this bit here is whether the tool will automatically 02:57.630 --> 02:59.220 be returned to the user, 02:59.220 --> 03:02.160 or whether the agent will take the output of the tool 03:02.160 --> 03:04.560 and then decide what to return to the user. 03:04.560 --> 03:06.570 Now how do you call a tool? 03:06.570 --> 03:08.280 Well, you can just do the tool.invoke, 03:08.280 --> 03:09.630 and you can see if I run this, 03:09.630 --> 03:11.487 it's gonna start using that tool 03:11.487 --> 03:14.910 and it gives us some information about digital marketing. 03:14.910 --> 03:17.790 Now you can see it's got a doc content chars_max, 03:17.790 --> 03:19.890 so we could also set this to 1,000, 03:19.890 --> 03:21.949 and then we're gonna get a lot more content back 03:21.949 --> 03:24.240 about that example. 03:24.240 --> 03:26.730 Now one thing that's often the case 03:26.730 --> 03:29.070 is you'll want to create your own tools. 03:29.070 --> 03:31.620 And your tools are gonna need a name, a description, 03:31.620 --> 03:34.530 and an arg schema, which is optional. 03:34.530 --> 03:38.070 The name is required and so the two optional ones 03:38.070 --> 03:39.979 are the description and the argschema, 03:39.979 --> 03:42.450 but it can be really nice to include those. 03:42.450 --> 03:44.940 Now the first way that you can easily create a tool 03:44.940 --> 03:46.530 is you could have a Python function, 03:46.530 --> 03:48.930 like where you gonna do a Google search? 03:48.930 --> 03:50.460 And we could take a query in here now. 03:50.460 --> 03:52.170 For now we're just returning LangChain, 03:52.170 --> 03:53.940 which is a type of string, 03:53.940 --> 03:57.180 and we wrap that specific Python function 03:57.180 --> 03:59.310 with this tool decorator. 03:59.310 --> 04:04.170 And then based on that, you will then from the doc string, 04:04.170 --> 04:06.570 it will add that into the description. 04:06.570 --> 04:07.860 So your doc strings, 04:07.860 --> 04:09.870 you want to make them very descriptive 04:09.870 --> 04:13.050 about what kind of tools your agents can use. 04:13.050 --> 04:16.410 The name of the tool comes from the Python function 04:16.410 --> 04:17.940 that you can see here. 04:17.940 --> 04:21.270 And the argschema or the .args in this case, 04:21.270 --> 04:23.460 get associated with the query. 04:23.460 --> 04:25.620 And this is basically the function signature. 04:25.620 --> 04:28.380 So you've got one input parameter here, which is a query, 04:28.380 --> 04:30.300 which is the type of string. 04:30.300 --> 04:33.960 Now you could also modify your inputs. 04:33.960 --> 04:36.099 And so you can see here we've got a didactic model, 04:36.099 --> 04:39.660 and we have one input which is a query type of string, 04:39.660 --> 04:42.060 but we've also added on a description 04:42.060 --> 04:44.220 to make it a bit more descriptive. 04:44.220 --> 04:46.440 And you can now see we've also changed the name 04:46.440 --> 04:48.120 of the tools to search-tool. 04:48.120 --> 04:50.160 And we've also said that this output 04:50.160 --> 04:53.520 should be immediately returned to the user, 04:53.520 --> 04:56.460 and not the agent if it's the final finish. 04:56.460 --> 05:01.020 So you can also make tools from the search function 05:01.020 --> 05:03.900 in terms of a structured tool .from function, 05:03.900 --> 05:06.840 where the function is here, you give it the name, 05:06.840 --> 05:10.224 and the description and you can even add on the argschema 05:10.224 --> 05:12.300 and all that good stuff. 05:12.300 --> 05:14.400 Now this shows you how to make tools, 05:14.400 --> 05:17.100 but let's see how do we add these onto agents? 05:17.100 --> 05:19.470 The easiest way to do that is you would have some function. 05:19.470 --> 05:21.600 So for example, we've got the get_word_length, 05:21.600 --> 05:25.710 we'd make our list of tools and we'd call that a tools list. 05:25.710 --> 05:27.630 And so what you end up here with is you've got 05:27.630 --> 05:32.430 that tools-list, which has a bunch of tools. 05:32.430 --> 05:34.410 Now once you have your bunch of tools, 05:34.410 --> 05:36.450 I personally find that the best way to use 05:36.450 --> 05:39.750 these LangChain agents is to write your own custom agents. 05:39.750 --> 05:41.790 Now it's quite easy to do that. 05:41.790 --> 05:43.650 There's a couple of different things that you'll need. 05:43.650 --> 05:46.500 So the first is you'll need a chat prompt template, 05:46.500 --> 05:48.240 and then you can have whatever you want 05:48.240 --> 05:49.770 in the system message. 05:49.770 --> 05:53.610 You need some form of input to take the original query. 05:53.610 --> 05:55.500 And then this is the most important bit 05:55.500 --> 05:56.970 that you have to have if you're building 05:56.970 --> 06:00.690 your own custom LangChain expression language agents. 06:00.690 --> 06:03.060 And basically this agent scratch pad 06:03.060 --> 06:07.980 is how LangChain will store all of those actions 06:07.980 --> 06:12.180 and observations so that it can use that in memory 06:12.180 --> 06:13.860 to then decide what to do next. 06:13.860 --> 06:16.290 'Cause remember there was that endless feedback loop 06:16.290 --> 06:19.440 of do an action, get an observation, 06:19.440 --> 06:22.170 decide on the next action, have we finished? 06:22.170 --> 06:23.490 Oh, we haven't finished. 06:23.490 --> 06:25.920 Let's go and run that next action again, 06:25.920 --> 06:28.530 until you end up finishing. 06:28.530 --> 06:30.540 So that agent scratch pad is really important 06:30.540 --> 06:33.118 because that stores all of the kind of intermediate steps 06:33.118 --> 06:34.710 that LangChain uses. 06:34.710 --> 06:37.140 And remember the input here is basically 06:37.140 --> 06:40.770 what you can use to specifically decide on the first query 06:40.770 --> 06:42.300 that hits this agent. 06:42.300 --> 06:45.030 Now after defining the chat prompt template, 06:45.030 --> 06:47.940 you'll then need to bind tools to your agent. 06:47.940 --> 06:50.970 And basically OpenAI have recently moved away 06:50.970 --> 06:52.260 from function calling, 06:52.260 --> 06:54.720 and they've moved to binding tools 06:54.720 --> 06:56.580 and using tools directly. 06:56.580 --> 06:59.250 So we use the .bind_tools function 06:59.250 --> 07:02.910 and that takes a list of tools and you've got a model 07:02.910 --> 07:06.390 and your model also needs to be capable of calling tools. 07:06.390 --> 07:08.040 Now once you've got your model 07:08.040 --> 07:10.860 and you've bound the tools to that specific model, 07:10.860 --> 07:12.810 it knows about those tools, 07:12.810 --> 07:15.030 but we still need to set up the formatting 07:15.030 --> 07:18.390 to easily convert the intermediate steps 07:18.390 --> 07:20.580 that are inside the agent scratch pad 07:20.580 --> 07:22.950 into something that OpenAI understands. 07:22.950 --> 07:25.380 So what's happening is these intermediate steps 07:25.380 --> 07:27.030 that the agents are scratch pad down, 07:27.030 --> 07:29.190 I've taken this action and run these tools, 07:29.190 --> 07:32.452 it's formatting those to the OpenAI tool messages 07:32.452 --> 07:35.730 and then after that gets put into the prompt, 07:35.730 --> 07:38.310 we then have the model that is with the tools 07:38.310 --> 07:41.940 and we also pass the tools outputs 07:41.940 --> 07:43.860 back into these intermediate steps 07:43.860 --> 07:47.280 so that the agent executor then can easily understand those 07:47.280 --> 07:49.410 and see what's happening with those. 07:49.410 --> 07:53.520 So after you've got your agent set up in an LCL chain, 07:53.520 --> 07:56.550 you would then add that onto the agent executor. 07:56.550 --> 07:58.020 You also pass the tools here 07:58.020 --> 08:00.660 and you can also put in verbose-True 08:00.660 --> 08:02.790 to make sure that comes across 08:02.790 --> 08:04.740 and you can see all the various steps. 08:04.740 --> 08:07.440 Now if I just .stream the output 08:07.440 --> 08:09.090 and you can see the input 08:09.090 --> 08:12.936 is gonna be mapped if we scroll back up to this input here 08:12.936 --> 08:16.650 in the user message, and we're saying something like, 08:16.650 --> 08:19.080 how many letters in the word data. 08:19.080 --> 08:21.360 And by doing agent_executed.stream, 08:21.360 --> 08:23.730 we can actually see all the different steps 08:23.730 --> 08:25.440 that happen inside of the agency. 08:25.440 --> 08:26.310 You've got the actions, 08:26.310 --> 08:28.560 it decided to take a tool agent action, 08:28.560 --> 08:30.990 and it used this get_word_length tool 08:30.990 --> 08:33.500 and we can also see what the final output is, 08:33.500 --> 08:35.820 so the word data is four letters 08:35.820 --> 08:38.340 and this is the messages that ended with. 08:38.340 --> 08:40.830 Now you wouldn't always stream these 08:40.830 --> 08:42.600 and the way that you would normally use these 08:42.600 --> 08:44.340 if you just wanted to get the final output 08:44.340 --> 08:46.530 is you would use that .invoke function. 08:46.530 --> 08:48.450 And by changing that to the .invoke, 08:48.450 --> 08:51.030 you'll see that we just get the final output, 08:51.030 --> 08:52.590 which will come through in just a second. 08:52.590 --> 08:55.800 So you get this input, how many letters is in the word data, 08:55.800 --> 08:58.590 and you get the output, the word data has four letters. 08:58.590 --> 09:00.210 Now that's great, 09:00.210 --> 09:03.570 but we don't currently have any memory inside of this agent. 09:03.570 --> 09:05.340 It can just use a tool, 09:05.340 --> 09:08.070 and it can use that to help solve a question. 09:08.070 --> 09:11.070 Now one thing that we can do to add memory in 09:11.070 --> 09:14.460 is we can add another messages placeholder, 09:14.460 --> 09:16.650 and you'll see we've got two messages placeholders. 09:16.650 --> 09:18.030 So one for the memory key, 09:18.030 --> 09:20.340 which is we've called chat history. 09:20.340 --> 09:22.320 And that chat history, that key name, 09:22.320 --> 09:25.050 the variable name here is very important. 09:25.050 --> 09:28.530 And you'll see this has also got our agent scratch pad. 09:28.530 --> 09:32.400 Now we create a Python list called Chat History, 09:32.400 --> 09:35.520 and we updated the LCL chain. 09:35.520 --> 09:39.990 So basically we're getting the chat history 09:39.990 --> 09:43.350 and then that's then being formatted into the prompt. 09:43.350 --> 09:45.540 So you've got both the chat history 09:45.540 --> 09:47.280 and the agent scratch pad 09:47.280 --> 09:48.660 that are now being associated 09:48.660 --> 09:51.120 with these message placeholders, yeah? 09:51.120 --> 09:54.240 So you can add as many message placeholders as you want, 09:54.240 --> 09:55.620 but you have to make sure that 09:55.620 --> 09:57.570 before that enters the prompt, 09:57.570 --> 09:59.430 you've got some output for those 09:59.430 --> 10:03.060 inside of your prompt template that goes into the prompt. 10:03.060 --> 10:03.990 Then it goes into the model 10:03.990 --> 10:06.510 and we pass back the agent tools agent 10:06.510 --> 10:07.980 using an output passer. 10:07.980 --> 10:10.290 The agent executor's exactly the same. 10:10.290 --> 10:11.850 But you'll see we've got this first input, 10:11.850 --> 10:13.860 how many letters in the word data, 10:13.860 --> 10:16.620 and we've now had to add on that chat history 10:16.620 --> 10:19.680 as the extra one, which is just a Python list. 10:19.680 --> 10:23.550 Now we then extend the chat history with a human message 10:23.550 --> 10:25.260 with the input one, 10:25.260 --> 10:29.160 and the AI's message that it came back with a result output. 10:29.160 --> 10:31.500 And then we can say, is that a real word? 10:31.500 --> 10:34.140 And so we can have a look. So firstly it's gonna figure out 10:34.140 --> 10:37.641 that it needs to invoke this tool called the get_word_length 10:37.641 --> 10:41.160 and it uses the input for that as data 10:41.160 --> 10:44.010 and then it basically work, 10:44.010 --> 10:45.300 this is the output it came back with, 10:45.300 --> 10:47.130 so the word data is four letters. 10:47.130 --> 10:48.210 And then after that, 10:48.210 --> 10:50.550 then it entered a new agent executed chain. 10:50.550 --> 10:52.980 But rather than deciding to invoke a tool, 10:52.980 --> 10:55.680 it decided to just reply with text. 10:55.680 --> 10:57.210 And that's quite an interesting thing, 10:57.210 --> 11:00.660 because you don't always have to invoke tools 11:00.660 --> 11:01.890 when using agents. 11:01.890 --> 11:04.800 They can use a mixture of just replying in text 11:04.800 --> 11:06.180 or using a tool. 11:06.180 --> 11:07.980 And in OpenAI tools, 11:07.980 --> 11:10.920 you can also run multiple tools in parallel, 11:10.920 --> 11:13.281 which can make agents incredibly versatile 11:13.281 --> 11:16.020 at not only being able to handle certain queries, 11:16.020 --> 11:18.570 but also being able to handle a mixture of queries. 11:18.570 --> 11:20.672 And we can see that the second response 11:20.672 --> 11:23.430 is saying yes, data is a real word. 11:23.430 --> 11:25.860 It refers to facts and statistics collected together 11:25.860 --> 11:26.940 for reference or analysis. 11:26.940 --> 11:29.370 So it is actually referring back to the fact 11:29.370 --> 11:33.630 that we asked it a question about what's the data? 11:33.630 --> 11:36.540 And so what's the count for that word. 11:36.540 --> 11:39.000 Now one thing that can be quite nice to do 11:39.000 --> 11:40.680 is obviously we've done this, 11:40.680 --> 11:43.830 but I just wanna show you another way of doing the memory. 11:43.830 --> 11:47.970 You could create, so when we're building these chains, 11:47.970 --> 11:50.070 they are something called a runnable. 11:50.070 --> 11:52.980 And a runnable is a something that can be ran. 11:52.980 --> 11:55.230 It's a LangChain expression 11:55.230 --> 11:57.990 that's basically compiled into a graph. 11:57.990 --> 11:59.430 Now if we go back down here, 11:59.430 --> 12:01.170 we can also do something like this, 12:01.170 --> 12:03.060 where we can create an empty store, 12:03.060 --> 12:05.460 which is gonna store our chat messages 12:05.460 --> 12:08.460 and we can do something with runnable with message history, 12:08.460 --> 12:10.770 and we can specify a session ID. 12:10.770 --> 12:12.030 And so what you'll see here is 12:12.030 --> 12:14.190 we've changed the variable name to history, 12:14.190 --> 12:16.230 just to show you how this works. 12:16.230 --> 12:18.660 The agent chain therefore has to update 12:18.660 --> 12:20.250 and it has to use history. 12:20.250 --> 12:22.620 The agent scratch pad stays the same. 12:22.620 --> 12:27.120 And what we've also got here is a get_session history 12:27.120 --> 12:28.980 that takes a session ID, 12:28.980 --> 12:32.550 looks inside of that Python dictionary, which is the store. 12:32.550 --> 12:34.080 And then after we've looked inside 12:34.080 --> 12:36.870 of that Python dictionary, we add on that session ID 12:36.870 --> 12:39.030 and we create this chat message history. 12:39.030 --> 12:42.480 And then we return the session ID. 12:42.480 --> 12:45.900 Now we can then wrap the agent executor 12:45.900 --> 12:47.280 with that Python function 12:47.280 --> 12:50.520 and the input messages key is gonna be the input. 12:50.520 --> 12:52.050 So remember that's our input here, 12:52.050 --> 12:55.650 and you can see that in the ACL inside of this agent chain. 12:55.650 --> 12:59.520 And the history messages key is the history. 12:59.520 --> 13:00.960 Now once we've done that, 13:00.960 --> 13:02.910 now what we can do is we can invoke, 13:02.910 --> 13:04.770 put our input, the history, 13:04.770 --> 13:07.620 set the config for the runnable config, 13:07.620 --> 13:09.769 and we say it's configurable with a session ID, 13:09.769 --> 13:10.860 some_session_ID. 13:10.860 --> 13:14.040 And what's happening now is the store, 13:14.040 --> 13:15.570 if we have a look at it, 13:15.570 --> 13:19.440 is actually gonna start to be populated with a session ID 13:19.440 --> 13:22.140 which has a chat message history attached to it. 13:22.140 --> 13:24.060 So this is an in-memory way 13:24.060 --> 13:27.180 of doing configurable chat message histories, 13:27.180 --> 13:30.960 which is being associated with this specific dictionary, 13:30.960 --> 13:33.540 this Python dictionary, this store Python dictionary. 13:33.540 --> 13:37.800 Now if we then say, I said, I've told it my name is James, 13:37.800 --> 13:40.080 and now we can also say what is my name? 13:40.080 --> 13:41.820 And we're using the same session ID, 13:41.820 --> 13:43.650 which we've called some_session_ID, 13:43.650 --> 13:46.770 and the agent executor hasn't used any tools yet. 13:46.770 --> 13:49.680 It's still got the ability to count with the word length, 13:49.680 --> 13:51.360 but it's decided to just reply in text. 13:51.360 --> 13:52.980 And I think that's an important point. 13:52.980 --> 13:54.480 And you can also see here 13:54.480 --> 13:56.460 that because we are using the same session ID 13:56.460 --> 13:58.080 in our Runnable config, 13:58.080 --> 14:01.440 we've then also been able to then get the LLM 14:01.440 --> 14:06.030 to store that chat history in a in-memory Python dictionary. 14:06.030 --> 14:08.220 And the output then says, your name is James. 14:08.220 --> 14:11.130 So it is able to remember who we are 14:11.130 --> 14:14.250 by using that with message history runnable. 14:14.250 --> 14:18.060 Now if I change the session ID to something different, 14:18.060 --> 14:20.280 you'll see that we then get a very different response 14:20.280 --> 14:21.113 from the agent, 14:21.113 --> 14:23.490 where it decides you haven't mentioned your name yet, 14:23.490 --> 14:26.190 please could you tell me your name, and I'll remember it. 14:26.190 --> 14:29.520 So hopefully this gives you a good introduction to agents. 14:29.520 --> 14:30.600 In future videos, 14:30.600 --> 14:32.970 we'll be exploring how you can build agents 14:32.970 --> 14:35.190 directly inside of things like line graph, 14:35.190 --> 14:38.010 and also how you could build multiple tools 14:38.010 --> 14:39.873 and looking at parallel tool usage.