WEBVTT 00:00.960 --> 00:02.070 -: In this video, we're gonna cover 00:02.070 --> 00:04.170 something called ConfigurableFields 00:04.170 --> 00:06.150 inside of LangChain Expression Language. 00:06.150 --> 00:08.610 Basically, configuration fields are a way for you 00:08.610 --> 00:11.790 to actually define specific bits of configuration 00:11.790 --> 00:15.360 that will be invoked or ran at runtime. 00:15.360 --> 00:17.340 Let's have a look at a couple of different examples 00:17.340 --> 00:18.780 so you can see how you can use this 00:18.780 --> 00:20.760 within LangChain Expression Language. 00:20.760 --> 00:23.010 We're firstly importing a couple of different packages, 00:23.010 --> 00:25.590 so langchain and langchain_openai. 00:25.590 --> 00:28.590 After doing that, we set up a very simple PromptTemplate 00:28.590 --> 00:29.640 and we've got our first thing, 00:29.640 --> 00:31.530 which is a ConfigurableField. 00:31.530 --> 00:33.480 Now, ConfigurableFields can be set 00:33.480 --> 00:36.330 on multiple different things such as the prompt. 00:36.330 --> 00:38.130 It can be done at the LLM. 00:38.130 --> 00:40.200 It can be done at lots of different types of things. 00:40.200 --> 00:42.780 These are the main two, though, prompts and LLMs. 00:42.780 --> 00:45.270 If we have a look here, we've got a ConfigurableField. 00:45.270 --> 00:48.690 We set up our chat model, which is a ChatOpenAI model, 00:48.690 --> 00:51.210 and we have this .configurable_fields. 00:51.210 --> 00:53.100 Notice how we set the temperature, 00:53.100 --> 00:55.320 which we know is a ConfigurableField. 00:55.320 --> 00:58.410 We know it's a field that we can access on ChatOpenAI. 00:58.410 --> 01:01.050 We create a ConfigurableField and we give it a name 01:01.050 --> 01:03.450 and a description and an id. 01:03.450 --> 01:05.490 Now, the id is the most important point 01:05.490 --> 01:08.190 because that's actually how you specifically change 01:08.190 --> 01:10.440 that field on whatever it is. 01:10.440 --> 01:12.630 So let's have an example and and see that. 01:12.630 --> 01:15.270 So we've got this model, and if we have a look now, 01:15.270 --> 01:16.620 if we just do the .invoke 01:16.620 --> 01:20.490 and we say pick a random number, it'll pick a random number, 01:20.490 --> 01:23.520 but it will be using the temperature as you can see here, 01:23.520 --> 01:25.050 which has been set to zero. 01:25.050 --> 01:28.740 Now, if we want to then modify that temperature at runtime, 01:28.740 --> 01:31.380 because we've set a ConfigurableField for the temperature, 01:31.380 --> 01:35.340 we can just do model.config, so it's model.with_config, 01:35.340 --> 01:37.830 and then you put a configurable object in there. 01:37.830 --> 01:41.220 Now, this is the same as or type of a RunnableConfig, 01:41.220 --> 01:42.330 and there's lots of different things 01:42.330 --> 01:44.400 you can add into a RunnableConfig. 01:44.400 --> 01:45.960 Now, in this scenario, 01:45.960 --> 01:49.770 notice we're referencing the llm_temperature, 01:49.770 --> 01:52.770 which is the same as we have for this id field 01:52.770 --> 01:54.690 when we create the ConfigurableField. 01:54.690 --> 01:57.510 And we're setting the temperature here to be 0.9, 01:57.510 --> 01:59.160 and then we are running the .invoke. 01:59.160 --> 02:02.910 So again, it's that model.with_config function 02:02.910 --> 02:05.160 that's then gonna allow you to change 02:05.160 --> 02:07.860 any ConfigurableField at runtime. 02:07.860 --> 02:09.480 You could also do this on a prompt. 02:09.480 --> 02:11.730 So for example, we can take a chat model, 02:11.730 --> 02:13.770 we can then specify a PromptTemplate, 02:13.770 --> 02:15.540 and after we're doing the PromptTemplate, 02:15.540 --> 02:18.750 we then have .configurable_alternatives. 02:18.750 --> 02:21.720 And this shows us that we could then have a different field. 02:21.720 --> 02:23.970 So for example, the ConfigurableField here 02:23.970 --> 02:25.530 is the idea of prompt. 02:25.530 --> 02:27.660 The default key will be a joke. 02:27.660 --> 02:31.053 So that basically means that will be the default LLM. 02:31.053 --> 02:33.480 This is the default key that will be used. 02:33.480 --> 02:36.090 However, you could also add a new option. 02:36.090 --> 02:39.000 So you can see here, for example, we have a poem 02:39.000 --> 02:42.240 and this says PromptTemplate.from_template, 02:42.240 --> 02:43.680 write a short poem. 02:43.680 --> 02:45.630 So what's really happening here 02:45.630 --> 02:47.550 is you've got this default prompt 02:47.550 --> 02:50.370 that's gonna be tell me a joke about topic, 02:50.370 --> 02:51.840 but we've also got the ability 02:51.840 --> 02:55.350 to have a different PromptTemplate here if we use the poem. 02:55.350 --> 02:56.310 So let's have a look at that. 02:56.310 --> 02:58.470 So we create our runnable chain. 02:58.470 --> 03:01.860 So you can see the prompt gets piped into the LLM, 03:01.860 --> 03:04.020 and then after that, then we can invoke that. 03:04.020 --> 03:07.290 So what's happening here is on this section above, 03:07.290 --> 03:09.300 you've got tell me a joke about topic 03:09.300 --> 03:10.800 and then we'll get the topic 03:10.800 --> 03:12.270 which will be injected with bears. 03:12.270 --> 03:13.590 So tell me a joke about bears 03:13.590 --> 03:14.760 and it'll say, why did the joke 03:14.760 --> 03:16.140 break up with his girlfriend? 03:16.140 --> 03:19.020 Because he couldn't bear the relationship any longer. 03:19.020 --> 03:21.900 But if we have a look now, we could also then say 03:21.900 --> 03:26.400 this chain.with_config, we'll pass in a configurable object, 03:26.400 --> 03:28.350 which is a prompt of poem, 03:28.350 --> 03:32.400 it's gonna change this to be write a short poem about topic. 03:32.400 --> 03:35.400 And what's happening here is then when we do the invoke, 03:35.400 --> 03:37.470 the topic is then going to be associated 03:37.470 --> 03:40.653 with write a short poem about bears. 03:41.820 --> 03:44.790 So you could easily change the different PromptTemplates 03:44.790 --> 03:49.233 based on having these extra keys that you put in here. 03:50.760 --> 03:53.370 Cool, you can also save your configurations. 03:53.370 --> 03:55.530 So if you do the chain.with_config, 03:55.530 --> 03:59.130 you can see you get a serializable JSON string, 03:59.130 --> 04:02.310 which you could also load up into an object as well. 04:02.310 --> 04:03.930 Just to have a look here as well, 04:03.930 --> 04:06.480 you can see that on LangChain's page, 04:06.480 --> 04:07.770 there's a couple of different other ways 04:07.770 --> 04:09.780 that you could use ConfigurableFields. 04:09.780 --> 04:12.360 So if you were using hubs, HubRunnables, 04:12.360 --> 04:13.770 you could use it for that. 04:13.770 --> 04:15.750 And also, there's a lot of different ways 04:15.750 --> 04:17.280 that you can optimize the LLM. 04:17.280 --> 04:18.750 So you could even say, 04:18.750 --> 04:22.050 I'm gonna make the LLM a RunnableConfigurableField. 04:22.050 --> 04:24.150 So the default one would be Anthropic, 04:24.150 --> 04:25.890 but then we'd also add on extra ones 04:25.890 --> 04:27.990 like OpenAI or GPT-4. 04:27.990 --> 04:30.900 And then you could specifically pick a chain, 04:30.900 --> 04:33.480 but then associate the LLM with different models. 04:33.480 --> 04:35.010 So your chain is actually able 04:35.010 --> 04:37.140 to run different models at runtime 04:37.140 --> 04:41.520 because you've made the model a ConfigurableField 04:41.520 --> 04:44.400 using that .configurable_alternatives method, 04:44.400 --> 04:47.520 and then adding on your ConfigurableFields. 04:47.520 --> 04:49.590 Cool, hopefully this gives you a good intro 04:49.590 --> 04:53.100 in how you can start using ConfigurableFields to optimize 04:53.100 --> 04:55.263 and do dynamic chains at runtime.