WEBVTT

00:00.400 --> 00:07.600
Let's discuss how guardrails AI uses prompts to generate precise and controlled outputs from a model.

00:08.760 --> 00:12.440
A guardrails prompt is not just a question or a command.

00:12.720 --> 00:19.400
It's a blueprint that guides the AI to produce outputs that meets our specific requirements.

00:20.160 --> 00:26.960
This blueprint combines instructions, parameters, and formatting guidelines, often in XML or JSON

00:26.960 --> 00:31.800
formats, to ensure the output aligns with our criteria.

00:31.840 --> 00:38.240
Imagine we need to generate a response to a billing inquiry from a patient who recently visited the

00:38.240 --> 00:38.960
hospital.

00:40.640 --> 00:43.040
Here is how the prompt structure looks like.

00:45.600 --> 00:47.360
We start with a clear directive.

00:47.360 --> 00:55.000
In this case, given the details of a patient's billing inquiry, please compile a JSON object that

00:55.000 --> 00:58.160
accurately reflects the requested billing information.

01:00.160 --> 01:04.040
This sets the stage informing the AI of the task at hand.

01:05.760 --> 01:11.960
Next, we introduce parameters that will be replaced at runtime with the actual input at hand.

01:12.400 --> 01:18.520
For instance, billing inquiry notes will contain the patient's inquiry details at the time of

01:22.540 --> 01:31.060
Finally, we include the boilerplate text like complete JSON suffix, which instructs the AI to format

01:31.060 --> 01:33.900
its outputs as a valid JSON object.

01:35.100 --> 01:42.700
This parameter will be replaced by the boilerplate schema definition, and the final prompt to be provided

01:42.700 --> 01:45.380
to LM looks like as shown here.

01:50.380 --> 01:57.460
The AI is tasked with extracting this information and presenting it in a structured JSON format, conforming

01:57.500 --> 02:00.020
to an XML schema provided here.

02:02.100 --> 02:09.060
For example, a patient's query about charges for consultation, diagnostic tests and treatments, along

02:09.060 --> 02:15.220
with insurance coverage details need to be meticulously transferred into a JSON object.

02:16.340 --> 02:24.220
The AI uses the provided XML schema to map the patient's inquiry to the correct JSON structure, ensuring

02:24.220 --> 02:32.420
every piece of information is accurately captured and formatted according to the specifications.

02:32.500 --> 02:39.100
This EDM process ensures that the output is not just accurate, but also precisely tailored to our specific

02:39.100 --> 02:42.940
needs.