WEBVTT 1 00:00:00.780 --> 00:00:05.040 Now besides uploading and handling images, 2 00:00:05.040 --> 00:00:09.090 which, again, must be supported by the model you're using 3 00:00:09.090 --> 00:00:09.990 to be an option, 4 00:00:09.990 --> 00:00:13.530 besides doing that, you can also attach other files, 5 00:00:13.530 --> 00:00:18.030 like for example, PDF, plain text or Word documents. 6 00:00:18.030 --> 00:00:19.590 Now, at least at the point of time 7 00:00:19.590 --> 00:00:20.850 where I'm recording this, 8 00:00:20.850 --> 00:00:22.410 an info pop-up opens 9 00:00:22.410 --> 00:00:25.380 if I choose the option to attach a file. 10 00:00:25.380 --> 00:00:26.820 And in this info pop-up, 11 00:00:26.820 --> 00:00:30.210 I get informed about the file types I can add, 12 00:00:30.210 --> 00:00:32.940 and that, at least right now when I'm recording this, 13 00:00:32.940 --> 00:00:36.090 I can upload up to five files 14 00:00:36.090 --> 00:00:39.900 with a maximum combined size of 30 megabytes. 15 00:00:39.900 --> 00:00:42.540 Now, of course, when you are viewing this video, 16 00:00:42.540 --> 00:00:44.490 you might no longer get this message. 17 00:00:44.490 --> 00:00:46.050 The limits also might have changed. 18 00:00:46.050 --> 00:00:49.620 But in general, when working with those local models, 19 00:00:49.620 --> 00:00:52.680 just slamming dozens of large files 20 00:00:52.680 --> 00:00:55.050 all into this one chat session 21 00:00:55.050 --> 00:00:57.120 typically won't be a good idea. 22 00:00:57.120 --> 00:00:59.640 If you wanna process multiple files, 23 00:00:59.640 --> 00:01:01.410 you should do it in chunk. 24 00:01:01.410 --> 00:01:03.960 The Context window size will also matter, 25 00:01:03.960 --> 00:01:06.660 but that's something I'll get back to later. 26 00:01:06.660 --> 00:01:09.810 In general, what LM Studio will try to do 27 00:01:09.810 --> 00:01:11.550 with any files you give it 28 00:01:11.550 --> 00:01:16.110 is it will try to load the content of this file 29 00:01:16.110 --> 00:01:19.890 into the chat history invisibly behind the scenes, 30 00:01:19.890 --> 00:01:22.350 as if you would have copy and pasted 31 00:01:22.350 --> 00:01:24.810 that content into the history. 32 00:01:24.810 --> 00:01:28.560 If that's not possible because the Context window, 33 00:01:28.560 --> 00:01:31.140 and again, I'll get back to this, is not sufficient, 34 00:01:31.140 --> 00:01:33.510 so if the model as it's currently configured 35 00:01:33.510 --> 00:01:36.570 would not be able to handle that amount of input, 36 00:01:36.570 --> 00:01:38.370 if that's not possible, 37 00:01:38.370 --> 00:01:41.580 LM Studio will actually try to split the content 38 00:01:41.580 --> 00:01:44.070 into smaller chunks for you 39 00:01:44.070 --> 00:01:46.890 and try to retrieve the relevant chunks 40 00:01:46.890 --> 00:01:50.430 based on the prompt you sent to the model. 41 00:01:50.430 --> 00:01:53.790 That's a technique called Retrieval Augmented Generation, 42 00:01:53.790 --> 00:01:57.870 and LM Studio will do it for you when uploading files. 43 00:01:57.870 --> 00:01:59.640 And again, the Context window part 44 00:01:59.640 --> 00:02:01.560 is something we'll explore later. 45 00:02:01.560 --> 00:02:03.090 Now here for this demo, 46 00:02:03.090 --> 00:02:07.053 I will upload a pretty small PDF document. 47 00:02:07.890 --> 00:02:12.360 It's just a very simple made up financial report PDF 48 00:02:12.360 --> 00:02:14.700 that includes a bunch of information 49 00:02:14.700 --> 00:02:18.270 about made up financials and numbers. 50 00:02:18.270 --> 00:02:19.103 But, of course, 51 00:02:19.103 --> 00:02:21.960 we can also use locally running large language models, 52 00:02:21.960 --> 00:02:25.320 like this one here, for summarizing documents like this. 53 00:02:25.320 --> 00:02:29.040 And indeed that is a task they're really, really good at. 54 00:02:29.040 --> 00:02:31.560 I will also say that the Gemma models, for example, 55 00:02:31.560 --> 00:02:34.200 do support quite large Context windows 56 00:02:34.200 --> 00:02:35.970 if you configure them appropriately, 57 00:02:35.970 --> 00:02:39.210 so they can handle quite a lot of information, 58 00:02:39.210 --> 00:02:41.790 and for example, summarize it for you. 59 00:02:41.790 --> 00:02:44.400 So for example, here I'll ask Gemma 3, 60 00:02:44.400 --> 00:02:46.620 the 12 billion parameters model again, 61 00:02:46.620 --> 00:02:47.880 to summarize this document 62 00:02:47.880 --> 00:02:51.210 and extract the key insights and financials for me, 63 00:02:51.210 --> 00:02:52.860 and that again, as I just mentioned, 64 00:02:52.860 --> 00:02:57.000 is one core use case for using large language models 65 00:02:57.000 --> 00:02:58.830 that are running locally on your system. 66 00:02:58.830 --> 00:03:00.360 They're really good at that, 67 00:03:00.360 --> 00:03:01.920 and of course, the huge advantage 68 00:03:01.920 --> 00:03:04.650 is that you don't need to share information like this, 69 00:03:04.650 --> 00:03:07.650 which potentially might, of course, be confidential 70 00:03:07.650 --> 00:03:10.020 or which you simply don't wanna share with everyone, 71 00:03:10.020 --> 00:03:12.270 with OpenAI or any other provider. 72 00:03:12.270 --> 00:03:15.870 Instead, you can perform tasks like this 73 00:03:15.870 --> 00:03:17.490 locally on your system 74 00:03:17.490 --> 00:03:19.890 by simply uploading documents to LM Studio 75 00:03:19.890 --> 00:03:21.330 and then working on them 76 00:03:21.330 --> 00:03:23.700 with your favorite large language model, 77 00:03:23.700 --> 00:03:25.593 in this case, the Gemma 3 model.