WEBVTT

00:00.520 --> 00:07.000
Let's take a few more examples to underscore how I guardrails can help with reliability, predictability,

00:07.200 --> 00:13.080
ethical use, and maintaining public trust in different domains.

00:13.080 --> 00:19.600
In the e-commerce service, customers interacting with an e-commerce site chatbot might ask for product

00:19.600 --> 00:24.560
recommendations or query order statuses.

00:24.560 --> 00:31.080
Input guardrails filter out irrelevant questions such as personal advice, ensuring the chatbot stays

00:31.080 --> 00:34.120
focused on e-commerce related inquiries.

00:34.840 --> 00:39.480
Output guardrails review the chatbots advice for accuracy.

00:39.560 --> 00:44.160
Ensuring recommendations or order information provided is correct and helpful.

00:44.840 --> 00:46.440
Preventing misinformation.

00:49.160 --> 00:56.760
In the field of mental health and wellness, users seeking support might input sensitive personal information

00:56.760 --> 00:58.280
or distress signals.

00:59.440 --> 01:06.120
Input guardrails filter out content that the app isn't designed to handle, such as immediate crisis

01:06.120 --> 01:11.680
intervention, directing users to human support if needed.

01:11.680 --> 01:18.320
Output girdles ensure responses are empathetic, accurate, and safely worded, avoiding language that

01:18.320 --> 01:20.920
could unintentionally cause distress or

01:24.930 --> 01:27.010
Talking about legal advice platforms.

01:27.010 --> 01:33.610
When users seeking legal advice, they input specific questions regarding their situation.

01:34.410 --> 01:41.330
Input guardrails ensure queries that fall outside the platform's legal expertise are flagged and not

01:41.330 --> 01:42.330
processed.

01:43.210 --> 01:50.290
Output guardrails check the legal advice generated for relevancy and general compliance, ensuring it

01:50.290 --> 02:01.850
doesn't provide incorrect or misleading information for the chatbot created for travel itinerary planning.

02:02.570 --> 02:08.930
Travelers input their preferences and constraints to receive personalized travel plans.

02:08.970 --> 02:15.490
Input garters filter out requests that are unfeasible or unsafe, like travel to conflict zones.

02:16.970 --> 02:23.770
Output guardrails ensure that the generated travel plans are practical, safe, and adhere to current

02:23.810 --> 02:25.170
travel advisories.

02:25.570 --> 02:30.810
Filtering out suggestions that don't meet the standards.

02:30.810 --> 02:36.690
In the next slides, we will look at two specific implementations of AI guardrails.

02:36.730 --> 02:39.770
Guardrails AI and Namo guardrails.

02:40.170 --> 02:47.970
We will see how they work, what they offer, and how they can be used with llms.