WEBVTT 00:00.880 --> 00:02.520 Hello everyone and welcome. 00:02.760 --> 00:06.600 In today's video we will learn about guard railing images with the prompt. 00:06.840 --> 00:09.960 So for that I'll first go ahead and create Guard Rail. 00:10.160 --> 00:13.120 Give the name as first link image Guard Rail. 00:13.360 --> 00:19.880 I'll skip the description since it's optional and the message for the blocked prompt would be. 00:19.880 --> 00:25.760 Sorry, the model cannot answer this question with image, and then I'll apply the same block message 00:25.760 --> 00:27.240 for responses as well. 00:27.800 --> 00:29.960 Go ahead, do next. 00:30.440 --> 00:34.960 Then we'll configure the content filters here for a harmful categories. 00:35.320 --> 00:42.120 We saw this in our previous video where we have hate insults, sexual violence, misconduct as a text. 00:42.120 --> 00:46.480 Amazon Bedrock has introduced image on the same category as well. 00:46.720 --> 00:53.920 So any of these categories, if they violate or have harmful categories with the image, it would guard 00:53.920 --> 00:55.560 rail against that as well. 00:56.240 --> 00:57.920 The strength is all the way high. 00:58.120 --> 01:01.040 And then I'll also go ahead and enable prompt attack. 01:01.440 --> 01:02.040 Go do. 01:02.040 --> 01:06.360 Next I'm going to skip all the other guard railing. 01:06.360 --> 01:12.950 That happens with different categories and different features because our focus is going to be mostly 01:12.950 --> 01:13.910 on images. 01:14.190 --> 01:17.790 And so I'll go ahead and then I'll skip profanity filter. 01:17.790 --> 01:23.670 I'll skip the word custom word phrases PII I'll skip the regex patterns. 01:24.110 --> 01:26.510 Also skip the contextual grounding check. 01:26.670 --> 01:31.990 And now all we have is enabled for harmful categories which is text and images. 01:33.430 --> 01:35.950 So now I'll go ahead and create the guardrail. 01:35.950 --> 01:42.070 So on this playground here I'll have to select a model in case the guardrail goes through. 01:42.550 --> 01:45.350 Then it is used to generate the response. 01:45.710 --> 01:52.630 So it's important to notice that to send an image you'll have to select the option that says use Apply 01:52.750 --> 01:54.270 Guardrails API. 01:54.630 --> 02:00.630 And for the input here you have prompts, uploaded text, images and content to be sent to the model. 02:00.870 --> 02:02.750 Now I'll go ahead and upload the image. 02:03.390 --> 02:05.510 So here I have selected this image. 02:05.630 --> 02:08.830 This is a text image that says example text. 02:08.870 --> 02:10.390 I'll just do a confirm. 02:10.710 --> 02:14.950 So now if you go ahead and check the check the image, it's a text image. 02:15.150 --> 02:17.590 And now I'm going to pass the content here. 02:17.750 --> 02:21.270 How can I find the taxes run this okay. 02:21.740 --> 02:25.300 so there was no problem with this input with the image. 02:25.300 --> 02:27.820 Let's go ahead and check the the trace here. 02:27.860 --> 02:30.820 No issues, no action were taken. 02:30.980 --> 02:32.260 This was all good. 02:32.540 --> 02:36.660 Now let's go back and use the very same image with a different text. 02:36.780 --> 02:40.020 So if you notice here you can upload it from your local. 02:40.260 --> 02:45.980 Or also you can use S3 bucket to upload the image I'm uploading from my local. 02:45.980 --> 02:47.420 So go ahead and confirm. 02:47.740 --> 02:53.380 This time I will ask question what are the loopholes for doing this right. 02:53.700 --> 02:56.820 And then we have the file here which is a text file. 02:57.060 --> 02:58.340 Go ahead and run this. 02:59.220 --> 02:59.900 All right. 03:00.420 --> 03:02.980 So now there was a guardrail action taken. 03:02.980 --> 03:04.060 It was intervened. 03:04.060 --> 03:08.460 And it said sorry the model cannot answer this question with image. 03:08.700 --> 03:10.100 It's with the trace here. 03:10.100 --> 03:16.500 And if you notice it categorized as a content filter, the test results were blocked. 03:16.500 --> 03:20.940 And the details here are the detected misconduct and the strength is high. 03:22.020 --> 03:28.700 So now Amazon Bedrock offers the ability to detect harmful intentions with image with the content. 03:29.020 --> 03:29.580 All right. 03:29.820 --> 03:30.940 Thank you so much. 03:30.940 --> 03:32.860 I'll see you in the next video.