WEBVTT 00:00.080 --> 00:02.520 So now let's run through the user input. 00:02.520 --> 00:04.840 That is not part of the disallowed topic. 00:04.880 --> 00:08.200 I'm asking a user input as how to play ping pong. 00:08.440 --> 00:10.840 Let's execute this and see what happens. 00:11.160 --> 00:18.200 So if you see here I have captured all the info logs and I get all the details about the execution of 00:18.200 --> 00:19.520 the Nemo runtime. 00:19.920 --> 00:22.520 Here you can see I got a response back. 00:23.360 --> 00:26.560 You can play ping pong by hitting the ball with the paddle. 00:27.000 --> 00:33.000 Should always try to hit the ball in a way that makes it difficult for other players to run it, or 00:33.040 --> 00:34.200 to return it. 00:34.560 --> 00:37.680 Here is all the details that happens on the runtime. 00:37.840 --> 00:44.560 I have captured those details and I removed things or execution that was not important. 00:44.880 --> 00:51.800 Let's go over the info logs and understand how Nemo executes the workflow on the back end. 00:52.680 --> 00:57.840 So from the current runtime architecture diagram, we saw that it executes three main stages. 00:58.040 --> 01:05.960 One is to generate canonical user message, decide next steps and execute them and generate bot utterances. 01:06.400 --> 01:10.210 Each of the above stage can involve one or more calls to LM. 01:10.730 --> 01:11.810 How does this happen? 01:12.130 --> 01:17.890 So in our case let's analyze our case here that says generate canonical user message. 01:18.650 --> 01:21.530 That's the first step that gets executed here. 01:21.530 --> 01:22.930 Let's zoom in a little bit. 01:23.290 --> 01:25.530 Utterance user action finished. 01:25.730 --> 01:27.530 How to play ping pong. 01:28.010 --> 01:35.370 That's the question that was asked at the very first phase of generating user input in our case here. 01:35.530 --> 01:43.530 If you see this is the sample conversation that we provided in the Config.yml, the runtime picks up 01:43.530 --> 01:48.490 this entire stack and it creates a conversation between user and the bot. 01:49.130 --> 01:52.050 There's another flow here how the user talks. 01:52.490 --> 01:58.650 This is nothing other than the calling specification that we have given for canonical user, canonical 01:58.650 --> 02:02.450 form and the user utterances user. 02:02.490 --> 02:04.330 How much do I have to boil pasta? 02:04.450 --> 02:06.810 And then that is ask about cooking. 02:07.130 --> 02:08.370 How do I rob a bank? 02:08.850 --> 02:13.250 That is criminal activity and user ask about criminal activity. 02:13.490 --> 02:16.730 All of this is the calling details that we have specified. 02:17.650 --> 02:21.250 Now this is the conversation between the user and the bot. 02:21.610 --> 02:24.290 So it combines the one that we had above. 02:24.570 --> 02:30.970 It says User Express greeting bot express greeting which is here user express greeting. 02:31.130 --> 02:32.610 Bot express greeting. 02:32.650 --> 02:35.250 It replies ask about capabilities. 02:36.090 --> 02:37.930 Then it appends the user input. 02:37.930 --> 02:39.130 How to play ping pong. 02:39.450 --> 02:46.450 This is then the entire stack is then passed to Lem here, which is in our case open air platform. 02:46.450 --> 02:52.930 And then understanding the entire conversation, the Lem comes up with the canonical form as ask about 02:52.930 --> 02:53.650 sport. 02:53.650 --> 02:59.850 It's very important to understand this here is that this entire conversation is for the bot or Lem. 02:59.890 --> 03:07.050 To understand that the format here ask about sport has to match, about ask about travel, ask about 03:07.050 --> 03:09.370 cooking all the canonical forms. 03:09.370 --> 03:16.530 It analyzes this, and it understands that how to play ping pong is a user utterance that falls under 03:16.530 --> 03:17.970 canonical form. 03:18.010 --> 03:19.300 Ask about sports. 03:20.220 --> 03:23.580 Then the conversation goes on about the details about it. 03:23.580 --> 03:24.700 How to play tennis. 03:25.180 --> 03:28.340 Tennis is a sport into which four players hit the ball. 03:28.540 --> 03:30.660 Then basketball and football. 03:30.940 --> 03:36.780 This keeps on going, but the canonical form for user intent is derived as ask about sport. 03:37.300 --> 03:43.940 Once that is determined, it moves on to the next step where it's the phase two for generating the next 03:43.940 --> 03:44.540 step. 03:44.740 --> 03:49.180 So the next step is like decide what the next steps are and execute them. 03:50.180 --> 03:52.940 So here if you see we provide the prompt. 03:53.820 --> 03:58.140 And then this is how the conversation between the user and the bot can go. 03:58.580 --> 04:03.820 This is no different than sample conversation here without the user utterances. 04:04.180 --> 04:09.180 These are very much just the canonical form that they're on time framework uses. 04:09.580 --> 04:13.140 Then again the other canonical forms how is the bot? 04:13.220 --> 04:14.860 This is how the bot thinks. 04:15.940 --> 04:20.700 User express about grid about cooking kernel activity. 04:21.100 --> 04:23.140 This is how the conversation is. 04:23.260 --> 04:26.150 And it understands that this is how the flow would go. 04:26.430 --> 04:32.630 And then the third one here is current conversation between the user and the bot is a mixture of or 04:32.670 --> 04:39.790 a union of user express greeting bot respond to capabilities and you should ask about sports. 04:40.310 --> 04:43.910 This was a canonical form that was generated by the Lem. 04:44.350 --> 04:46.150 We add that or will not be. 04:46.150 --> 04:52.910 The framework adds that to the to the existing conversation, and then passes on to the OpenAI DaVinci 04:52.950 --> 04:53.510 model. 04:54.750 --> 04:56.070 And the bot response. 04:56.070 --> 04:59.430 Here is bot response for sports. 04:59.630 --> 05:03.190 That's the next step that it generated for the flow. 05:03.550 --> 05:09.110 So the first one is user ask about sport and then the next one is bot respond for sports. 05:09.430 --> 05:15.030 This is after understanding how the interaction is between the user and the bot, and then it also goes 05:15.070 --> 05:20.630 ahead and creates the own its own line of kolong flows that is printed here. 05:21.510 --> 05:27.230 Then it moves on to the third phase for generating the bot message or generating the bot utterances. 05:27.670 --> 05:34.910 This is again a very similar flow where we provide the prompt and then hear the conversation between 05:34.910 --> 05:36.150 the user and the bot. 05:36.350 --> 05:40.750 Is the config.yml file with the sample conversation? 05:42.230 --> 05:45.910 Then this is the current conversation between the user and the bot. 05:46.750 --> 05:51.350 Now it knows that it's a conversation that is ongoing between the bot and the user. 05:51.670 --> 05:54.150 So here express greetings. 05:54.270 --> 05:57.510 More express greetings capabilities. 05:57.670 --> 05:59.870 User ask about capabilities. 06:00.310 --> 06:03.790 But then here is the Kallang flow that that was generated. 06:04.110 --> 06:09.150 One is the user's asking or the user utterance is how to play ping pong. 06:09.430 --> 06:12.190 This is categorized as asking about sport. 06:12.590 --> 06:14.910 And then bot responses for sport. 06:15.110 --> 06:21.030 So it has to generate the response for sport here which is passed to the open AI platform. 06:21.390 --> 06:23.630 And here we get the response back. 06:23.990 --> 06:27.750 Ping pong is a sport played by two workforce people. 06:27.990 --> 06:31.150 And the details that come back from the interaction. 06:31.390 --> 06:37.310 So this is how the three stages of the Nemo guardrail runtime is executed behind the scenes.