WEBVTT

00:00.840 --> 00:02.640
Hello everyone and welcome.

00:03.000 --> 00:07.080
In today's session we will learn about runtime execution of agents.

00:07.400 --> 00:10.720
The first step here involves pre-processing of the data.

00:10.960 --> 00:13.280
For that we need user input.

00:13.760 --> 00:20.120
Agents will then fetch any augmented prompt associated with the data or user input.

00:20.920 --> 00:25.800
It then looks for any conversation history associated with the user input.

00:26.160 --> 00:28.120
These two steps are optional.

00:28.560 --> 00:35.280
An application may or may not have an augmented prompt and a conversation history associated with a

00:35.280 --> 00:36.320
user input.

00:36.920 --> 00:45.080
Then it manages to contextualize and categorize the user input, and can be used to validate the input.

00:45.440 --> 00:50.040
Here, the category might involve whether the input is not harmful.

00:50.960 --> 00:57.400
It validates whether the input matches the required categories that a use case has, and then it moves

00:57.400 --> 01:00.600
on to the next phase, that is, the orchestration.

01:00.960 --> 01:07.400
The agent interprets the input with a foundation model, and generates a rationale that lays out the

01:07.400 --> 01:09.920
logic for the next step it should take.

01:10.280 --> 01:17.740
It generates and parses the response from the foundation model and predicts which action it should take

01:18.420 --> 01:21.860
if the agent determines that it has to invoke an action.

01:22.220 --> 01:27.580
The agent sends the parameters determined from the user input to a lambda function.

01:27.820 --> 01:34.900
If it determines that it has to invoke an associated knowledge base to retrieve additional context,

01:35.260 --> 01:41.940
then it invokes the passes the query to the knowledge base and gets the response back.

01:42.580 --> 01:50.100
The agent then generates an output known as the observation, from invoking an action, or summarizing

01:50.100 --> 01:52.060
the results from the knowledge base.

01:52.460 --> 01:59.940
This loop continues until the agent returns the response to the user, or until it needs to prompt the

01:59.940 --> 02:02.060
user for extra information.

02:02.980 --> 02:07.340
This is how the entire runtime execution happens for the agents.

02:07.660 --> 02:13.740
In the post-processing step, the agent formats the final response to the user and it's done.

02:14.140 --> 02:21.980
Now let's understand and run a hands on exercise in our next video to see all of this in action.

02:22.140 --> 02:22.940
Thank you.

02:22.980 --> 02:24.980
I'll see you in the next video.