WEBVTT

00:00.760 --> 00:05.000
Let's analyze single calls from the info log here.

00:05.320 --> 00:06.720
Let me zoom in a little bit.

00:07.000 --> 00:10.440
So for the user or the utterance user action finished.

00:10.800 --> 00:17.240
When I asked how to play ping pong, the runtime created an event with all three phases in one call.

00:17.720 --> 00:23.960
Here is where the improvisation happens and the framework optimizes the workflow by calling all the

00:23.960 --> 00:26.840
three phases in one call.

00:27.600 --> 00:30.000
Here is the prompt that we have specified.

00:30.400 --> 00:34.280
And here is the sample conversation between the user and the bot.

00:34.680 --> 00:39.320
We got this from the sample conversation in the config file.

00:39.640 --> 00:44.960
Here is where the user generates the next step for each user message.

00:45.080 --> 00:48.760
Generate the next steps and finish with the bot message.

00:49.000 --> 00:54.680
Here is where we instruct LA to generate next step and finish with this bot message.

00:54.880 --> 01:00.440
We provide this canonical form and user utterances from the disallowed topic file.

01:00.760 --> 01:08.840
And then we again instruct LM on the next line, generate bot message to relate it to express capabilities.

01:09.200 --> 01:13.320
And here are all the different conversations from the sample conversation.

01:14.040 --> 01:17.680
And at the end we append user question or query.

01:17.920 --> 01:19.760
That said, how to play ping pong.

01:20.040 --> 01:21.400
Pass this to LM.

01:21.600 --> 01:28.070
And then we got a response back here, clearly says what was the canonical form for the user input?

01:28.430 --> 01:31.190
The question was about ask about sports.

01:31.470 --> 01:36.270
Now it also responded bot respond for the sports gives an answer here.

01:37.110 --> 01:38.670
Ping pong is a sport.

01:38.950 --> 01:45.110
Just like that, we have some more canonical forms and bot responses attached to the response.

01:45.510 --> 01:46.670
How to play tennis.

01:46.670 --> 01:48.470
And this is ask about sport.

01:48.630 --> 01:54.670
And what response for the sports saying tennis is a sport that can be played individually.

01:55.110 --> 02:03.630
So now if you see trace down, the LM call was made in 1.78 seconds and canonical form for user intent

02:03.630 --> 02:09.670
was asked about sport and the canonical form for the bot intent was response for sports.

02:10.550 --> 02:13.910
And then the generate bot message was created as well.

02:14.270 --> 02:21.190
So that all the three steps were executed in one call in 1.83 seconds altogether.

02:21.630 --> 02:27.430
And I got the response back with the ping pong is a sport to which four players hit a line.

02:27.430 --> 02:29.990
Lightweight ball back in or not.

02:30.030 --> 02:30.950
All of that.

02:31.270 --> 02:37.550
So the goal here is that there was a lot of improvisation that happened with one call to LM having all

02:37.590 --> 02:38.870
the three stages.

02:39.150 --> 02:44.710
And it was able to execute the runtime more effectively by doing just one LM call.