WEBVTT 1 00:00:00.150 --> 00:00:01.320 Hey there, Ethan here. 2 00:00:01.320 --> 00:00:03.420 And I want in this video 3 00:00:03.420 --> 00:00:06.210 to dive a bit deeper about the implementation 4 00:00:06.210 --> 00:00:07.650 of structured output. 5 00:00:07.650 --> 00:00:09.240 So as we saw in the course, 6 00:00:09.240 --> 00:00:11.790 structured output allows agents 7 00:00:11.790 --> 00:00:16.410 to return data in specific, predictable format. 8 00:00:16.410 --> 00:00:20.880 So instead of getting a raw-text answer, 9 00:00:20.880 --> 00:00:24.000 we can get a Pydantic object, 10 00:00:24.000 --> 00:00:26.040 we can get a JSON response, 11 00:00:26.040 --> 00:00:27.900 or a dataclass, 12 00:00:27.900 --> 00:00:31.410 and something which is downstreamable to our application 13 00:00:31.410 --> 00:00:32.910 so we can use later. 14 00:00:32.910 --> 00:00:35.280 Now, we saw how we can use it 15 00:00:35.280 --> 00:00:38.670 in the create a agent function, right? 16 00:00:38.670 --> 00:00:42.450 Now, there are a bunch of implementations 17 00:00:42.450 --> 00:00:44.220 to structured output. 18 00:00:44.220 --> 00:00:46.770 So using structured output 19 00:00:46.770 --> 00:00:50.880 has two different types of implementation. 20 00:00:50.880 --> 00:00:53.700 The first one is tool strategy, 21 00:00:53.700 --> 00:00:57.240 which is going to use tool calling under the hood, 22 00:00:57.240 --> 00:01:00.420 and the second is called provider strategy. 23 00:01:00.420 --> 00:01:02.010 And this is going to be using 24 00:01:02.010 --> 00:01:07.010 the native model structured output capabilities. 25 00:01:07.230 --> 00:01:11.880 Now, LangChain is going to choose by default 26 00:01:11.880 --> 00:01:13.890 to use the provider strategy, 27 00:01:13.890 --> 00:01:15.930 unless we specify otherwise. 28 00:01:15.930 --> 00:01:19.440 So let's start talking about the provider strategy. 29 00:01:19.440 --> 00:01:21.900 So most top-tier models, 30 00:01:21.900 --> 00:01:26.160 they offer support for structured output natively. 31 00:01:26.160 --> 00:01:29.370 So this means we can use their own APIs 32 00:01:29.370 --> 00:01:33.060 and use a structured output argument, 33 00:01:33.060 --> 00:01:35.580 and they're going to do all the heavy loading 34 00:01:35.580 --> 00:01:38.370 and they're going to make the response 35 00:01:38.370 --> 00:01:40.020 be in the format that we want. 36 00:01:40.020 --> 00:01:42.930 And we can see that it's going to receive 37 00:01:42.930 --> 00:01:44.250 a bunch of schema types. 38 00:01:44.250 --> 00:01:46.757 So it can be a Pydantic model, a dataclass, 39 00:01:46.757 --> 00:01:49.500 a TypedDict and JSON scheme. 40 00:01:49.500 --> 00:01:52.290 We can see it right over here all the examples. 41 00:01:52.290 --> 00:01:54.480 This is for a Pydantic model, 42 00:01:54.480 --> 00:01:56.280 this is for a dataclass, 43 00:01:56.280 --> 00:01:57.660 this is for a TypedDict, 44 00:01:57.660 --> 00:01:59.940 and this is for a JOSN scheme. 45 00:01:59.940 --> 00:02:01.290 And by default, 46 00:02:01.290 --> 00:02:04.230 if the model we're going to be using in the create agent 47 00:02:04.230 --> 00:02:07.590 is going to support structured output, 48 00:02:07.590 --> 00:02:09.750 then LangChain is going to be using that, 49 00:02:09.750 --> 00:02:10.920 because by doing it, 50 00:02:10.920 --> 00:02:13.110 they're shifting all of the responsibility 51 00:02:13.110 --> 00:02:14.760 to the model provider 52 00:02:14.760 --> 00:02:16.470 to return us a good answer. 53 00:02:16.470 --> 00:02:20.325 And this basically means that if we get a wrong answer, 54 00:02:20.325 --> 00:02:23.490 (laughing) we can talk to the vendor and not LangChain. 55 00:02:23.490 --> 00:02:28.490 Now, some models do not offer this structured output API. 56 00:02:29.250 --> 00:02:32.370 In case the model and all of the top-tier models 57 00:02:32.370 --> 00:02:33.990 supports tool calling, 58 00:02:33.990 --> 00:02:35.730 we have a workaround to do it, 59 00:02:35.730 --> 00:02:38.070 and this is actually how everything evolved. 60 00:02:38.070 --> 00:02:40.860 So we can actually use tool calling 61 00:02:40.860 --> 00:02:43.020 and LangChain under the hood 62 00:02:43.020 --> 00:02:44.400 uses tool calling 63 00:02:44.400 --> 00:02:47.190 with the schema of the object we want 64 00:02:47.190 --> 00:02:51.900 in providing one single tool to the LLM 65 00:02:51.900 --> 00:02:53.610 and saying to the LLM 66 00:02:53.610 --> 00:02:55.770 they always need to choose that tool. 67 00:02:55.770 --> 00:02:56.910 And by doing that, 68 00:02:56.910 --> 00:02:58.830 we are forcing the LLM 69 00:02:58.830 --> 00:03:01.590 to enforce the schema that we send it. 70 00:03:01.590 --> 00:03:04.080 So this is a nice workaround 71 00:03:04.080 --> 00:03:05.910 to get structured output here. 72 00:03:05.910 --> 00:03:08.820 So LangChain now implement this for us. 73 00:03:08.820 --> 00:03:09.720 And in this case, 74 00:03:09.720 --> 00:03:12.241 it also supports all the Pydantic models, 75 00:03:12.241 --> 00:03:15.573 dataclass, TypedDict and JSON scheme.