WEBVTT 00:00:00.000 --> 00:00:02.720 Asking an AI to trace a single code base request 00:00:02.720 --> 00:00:07.200 cost 200 ,000 tokens. Then by giving the AI a 00:00:07.200 --> 00:00:10.460 simple map first, that cost dropped to 80 ,000 00:00:10.460 --> 00:00:13.279 tokens. Yeah, that is a massive drop and wasted 00:00:13.279 --> 00:00:15.380 computational effort. I mean, we're literally 00:00:15.380 --> 00:00:17.940 changing the starting line here. Welcome to the 00:00:17.940 --> 00:00:19.600 Deep Dive. I'm really glad you are joining us 00:00:19.600 --> 00:00:21.679 today. We have a pretty fascinating shift to 00:00:21.679 --> 00:00:24.519 talk about. We really do. You know, I still wrestle 00:00:24.519 --> 00:00:27.519 with prompt drift myself, especially when I ask 00:00:27.519 --> 00:00:30.559 an AI to analyze a massive project and it just 00:00:30.559 --> 00:00:33.200 gets hopelessly lost. Oh, absolutely. It's incredibly 00:00:33.200 --> 00:00:35.920 frustrating when it just spins its wheels. Exactly. 00:00:36.179 --> 00:00:38.899 So today we are exploring Graphify. It's an open 00:00:38.899 --> 00:00:41.640 source tool that basically gives cloud code a 00:00:41.640 --> 00:00:43.850 reusable memory. Which is something developers 00:00:43.850 --> 00:00:46.170 have been desperately asking for. Right. We'll 00:00:46.170 --> 00:00:47.789 cover how it builds a map of your repository. 00:00:48.090 --> 00:00:49.969 We'll look at the three -pass system it uses 00:00:49.969 --> 00:00:52.549 to understand your files. And we'll see how it 00:00:52.549 --> 00:00:54.950 fundamentally changes the way AI interacts with 00:00:54.950 --> 00:00:58.450 code bases. Okay, let's unpack this. We know 00:00:58.450 --> 00:01:01.350 AI gets lost in big projects. Why does that actually 00:01:01.350 --> 00:01:03.750 happen? Well, mainly because it totally lacks 00:01:03.750 --> 00:01:06.230 inherent structure when it begins. Right. Normally, 00:01:06.370 --> 00:01:09.390 cloud code searches file by file. It spawns these 00:01:09.390 --> 00:01:12.230 little explore agents, opens folders and just, 00:01:12.310 --> 00:01:14.870 you know, builds context from zero. It's kind 00:01:14.870 --> 00:01:16.849 of like trying to navigate a new city by walking 00:01:16.849 --> 00:01:18.769 down every random alleyway to see where it goes. 00:01:18.849 --> 00:01:21.349 Yeah, exactly. It's chaos. Versus just looking 00:01:21.349 --> 00:01:23.329 at a subway map before you leave the house. Right. 00:01:23.739 --> 00:01:26.180 And GraphDefy fixes this by turning the repo 00:01:26.180 --> 00:01:28.900 into a knowledge graph first. It builds that 00:01:28.900 --> 00:01:31.900 subway map. Why does crawling a repository manually 00:01:31.900 --> 00:01:35.079 eat up so many tokens? Because the AI is wasting 00:01:35.079 --> 00:01:37.040 context windows just trying to figure out the 00:01:37.040 --> 00:01:39.120 shape of the project before it even begins to 00:01:39.120 --> 00:01:41.219 answer the actual prompt. So it builds a map 00:01:41.219 --> 00:01:43.760 first, saving the AI from wandering blindly. 00:01:44.079 --> 00:01:46.260 Yes, exactly. And the way it accomplishes that 00:01:46.260 --> 00:01:49.019 is really methodical. It uses a three -pass approach, 00:01:49.159 --> 00:01:51.640 right? How does it start? Pass 1 focuses purely 00:01:51.640 --> 00:01:54.040 on the code. It uses a tool called TreeSitter 00:01:54.040 --> 00:01:57.299 to read the code structure locally. Classes, 00:01:57.299 --> 00:02:00.019 functions, imports. And this pass is completely 00:02:00.019 --> 00:02:02.140 deterministic. Wait, let me stop you right there. 00:02:02.239 --> 00:02:04.939 What does deterministic mean here? It follows 00:02:04.939 --> 00:02:08.159 strict logical rules without any AI guessing 00:02:08.159 --> 00:02:10.620 involved. Got it. So it just reads the actual 00:02:10.620 --> 00:02:14.120 mathematical connections. Right. No LLM hallucination 00:02:14.120 --> 00:02:19.159 yet. Beat. Then pass to handles audio and video. 00:02:19.319 --> 00:02:22.539 It uses faster whisper to transcribe media files. 00:02:22.740 --> 00:02:26.439 Like internal demo videos or meeting clips. Yeah, 00:02:26.479 --> 00:02:28.400 exactly. Those transcripts are added directly 00:02:28.400 --> 00:02:30.900 to the map. Wow, that is incredibly useful for 00:02:30.900 --> 00:02:32.719 capturing raw institutional knowledge. It's a 00:02:32.719 --> 00:02:36.110 total game changer. Then we hit Pass 3. This 00:02:36.110 --> 00:02:38.469 is where the large language model steps in for 00:02:38.469 --> 00:02:40.810 documents and images. Like PDFs and markdown 00:02:40.810 --> 00:02:43.169 files. What's fascinating here is how it blends 00:02:43.169 --> 00:02:46.409 rigid logic with AI interpretation. It understands 00:02:46.409 --> 00:02:49.250 the business logic in a design PDF and connects 00:02:49.250 --> 00:02:51.530 it back to the code from Pass 1. If Pass 1 is 00:02:51.530 --> 00:02:53.789 free and local, why do we even need the LLM in 00:02:53.789 --> 00:02:57.099 Pass 3? Code is rigid and easy to map, but human 00:02:57.099 --> 00:03:00.520 context, like design PDFs, requires AI to interpret 00:03:00.520 --> 00:03:02.780 the meaning. It reads code rules, exactly, then 00:03:02.780 --> 00:03:05.800 uses AI for the messy human stuff. Exactly. It 00:03:05.800 --> 00:03:08.580 creates a complete picture. So once that map 00:03:08.580 --> 00:03:11.020 is built... What does this web of data actually 00:03:11.020 --> 00:03:13.259 look like in practice? Right. It breaks down 00:03:13.259 --> 00:03:16.539 into three core parts. Nodes, edges, and communities. 00:03:16.860 --> 00:03:19.080 Let's define those for everyone. Nodes are the 00:03:19.080 --> 00:03:22.360 distinct entities, a file, a function, or a document. 00:03:22.620 --> 00:03:25.099 Okay, isolated pieces of information. Mm -hmm. 00:03:25.659 --> 00:03:28.740 Edges are the lines connecting those individual 00:03:28.740 --> 00:03:31.319 nodes, the actual functional relationship. On 00:03:31.319 --> 00:03:34.139 communities. Communities are organic clusters 00:03:34.139 --> 00:03:36.979 of highly related nodes, like neighborhoods in 00:03:36.979 --> 00:03:39.819 our subway map. The developers ran a really great 00:03:39.819 --> 00:03:42.979 demo of this. on the Open Design Repository. 00:03:43.039 --> 00:03:45.219 Yeah, it's a huge open source code base. Perfect 00:03:45.219 --> 00:03:48.800 stress test. Graphify scanned 203 files in about 00:03:48.800 --> 00:03:52.400 six minutes. And it created 197 distinct nodes 00:03:52.400 --> 00:03:57.960 plus over 3 ,400 edges. That is a massive web 00:03:57.960 --> 00:04:00.819 of dependencies. Yeah. And it found 109 communities, 00:04:01.000 --> 00:04:04.020 right? Yeah. It highlights surprising connections. 00:04:04.199 --> 00:04:06.199 It even identifies these things called God nodes. 00:04:06.580 --> 00:04:09.699 Ah, the highly connected critical parts of the 00:04:09.699 --> 00:04:13.379 repo. Whoa. Yeah. Beat. Imagine scaling that. 00:04:13.539 --> 00:04:16.060 Imagine seeing God nodes in a code base you've 00:04:16.060 --> 00:04:18.180 worked on for years but never truly understood. 00:04:18.439 --> 00:04:22.100 Beat. It completely exposes the hidden, fragile 00:04:22.100 --> 00:04:24.819 architecture of your project. What practical 00:04:24.819 --> 00:04:27.279 value does knowing the communities give to the 00:04:27.279 --> 00:04:30.500 AI? It allows clawed code to isolate the specific 00:04:30.500 --> 00:04:33.019 neighborhood of code it needs to solve a problem, 00:04:33.199 --> 00:04:36.689 ignoring irrelevant files entirely. Nodes are 00:04:36.689 --> 00:04:39.529 things, edges are connections, communities are 00:04:39.529 --> 00:04:42.149 specific code neighborhoods. That perfectly summarizes 00:04:42.149 --> 00:04:44.819 it. It's why it beats traditional search. Since 00:04:44.819 --> 00:04:46.899 we're talking about memory, listeners are probably 00:04:46.899 --> 00:04:49.139 wondering how this differs from GraphArea. They 00:04:49.139 --> 00:04:51.319 do sound similar, but the core difference is 00:04:51.319 --> 00:04:53.939 that Graphify does not use embeddings. Let's 00:04:53.939 --> 00:04:56.160 clarify that. What are embeddings? Translating 00:04:56.160 --> 00:04:59.160 text into numbers so AI can find similar underlying 00:04:59.160 --> 00:05:01.339 meanings. Here's where it gets really interesting. 00:05:01.819 --> 00:05:05.620 Graphify maps actual existing structural relationships. 00:05:06.399 --> 00:05:09.339 GraphArea searches giant unstructured document 00:05:09.339 --> 00:05:12.790 collections. Exactly. GraphRag finds things that 00:05:12.790 --> 00:05:15.990 sound related. Graphify reads exact, deterministic 00:05:15.990 --> 00:05:18.970 relationships that already exist, like code imports. 00:05:19.310 --> 00:05:22.029 It's basically RG Lite for repositories. Yeah, 00:05:22.129 --> 00:05:25.709 mapping reality, not probability. So if I have 00:05:25.709 --> 00:05:30.069 10 ,000 messy policy PDFs, I shouldn't use Graphify. 00:05:30.250 --> 00:05:33.509 Correct. GraphRag is for massive, unstructured 00:05:33.509 --> 00:05:36.790 text. Graphify is for structured systems where 00:05:36.790 --> 00:05:39.129 you want to know exactly how A connects to B. 00:05:39.600 --> 00:05:42.339 Graphify map structured systems. GraphRag searches 00:05:42.339 --> 00:05:45.319 giant piles of loose documents. Spot on. You 00:05:45.319 --> 00:05:47.639 have to choose the right tool for the job. Let's 00:05:47.639 --> 00:05:50.339 get back to the practical payoff. Time, money, 00:05:50.420 --> 00:05:52.220 and keeping this map alive. The token savings 00:05:52.220 --> 00:05:55.399 are very real here. In that test, asking Claude 00:05:55.399 --> 00:05:57.800 to trace a design request, the standard run cost 00:05:57.800 --> 00:06:01.800 200 ,000 tokens. And 150 ,000 of those were wasted 00:06:01.800 --> 00:06:04.160 just on explore agents. The AI was just frantically 00:06:04.160 --> 00:06:06.540 bumping into walls. Right. But the Graphify run 00:06:06.540 --> 00:06:09.279 took only 80 ,000 tokens total. That is about 00:06:09.279 --> 00:06:11.959 40 % of the original cost. Now, I do want to 00:06:11.959 --> 00:06:14.759 warn listeners not to expect the massive 70 times 00:06:14.759 --> 00:06:17.259 token savings hype you might see online. Yeah, 00:06:17.300 --> 00:06:19.540 the internet loves to exaggerate. Always. But 00:06:19.540 --> 00:06:22.420 40 % is still massive for big projects. And setup 00:06:22.420 --> 00:06:24.959 is incredibly easy. You just ask Cloud Code to 00:06:24.959 --> 00:06:27.800 install it. You just run slash graphify dot. 00:06:27.959 --> 00:06:31.040 And there's this really cool dash dash obsidian 00:06:31.040 --> 00:06:33.600 flag. Which turns the map into an obsidian vault 00:06:33.600 --> 00:06:36.540 for personal notes. It's so fun to visually explore 00:06:36.540 --> 00:06:39.519 your code base that way. But code changes daily. 00:06:39.800 --> 00:06:43.240 How do we keep this a living repo? You run Graphify 00:06:43.240 --> 00:06:46.360 hook install. It rebuilds the map automatically 00:06:46.360 --> 00:06:49.199 on every single commit. If my team pushes 50 00:06:49.199 --> 00:06:52.180 commits a day, won't updating this map get incredibly 00:06:52.180 --> 00:06:54.540 expensive? No, because the structural code updates 00:06:54.540 --> 00:06:57.319 don't use the LLM API. They just use the local 00:06:57.319 --> 00:07:00.139 free parsing tool. The hook updates the map locally 00:07:00.139 --> 00:07:02.360 and for free on every commit. It's completely 00:07:02.360 --> 00:07:06.240 sustainable. So what does this all mean? Graphify 00:07:06.240 --> 00:07:08.839 isn't just a cost -saving trick. No, not at all. 00:07:08.920 --> 00:07:10.860 If we connect this to the bigger picture, it's 00:07:10.860 --> 00:07:12.800 a fundamental shift in how we work with these 00:07:12.800 --> 00:07:15.839 tools. We are moving from treating an AI like 00:07:15.839 --> 00:07:18.120 a search engine that has to rediscover your project 00:07:18.120 --> 00:07:20.579 every morning to giving it a permanent living 00:07:20.579 --> 00:07:23.180 memory. This raises an important question about 00:07:23.180 --> 00:07:26.120 the future of development. It gives the AI genuine 00:07:26.120 --> 00:07:29.699 situational awareness. Think about the obsidian 00:07:29.699 --> 00:07:33.449 export feature. If Graphify can instantly map 00:07:33.449 --> 00:07:36.529 the complex logic of a massive software project 00:07:36.529 --> 00:07:38.990 and turn it into a visual vault, actually wait. 00:07:39.209 --> 00:07:41.970 What happens when we start pointing these deterministic 00:07:41.970 --> 00:07:44.170 mapping tools at our own personal hard drives, 00:07:44.269 --> 00:07:46.009 our personal notes, or our scattered thoughts? 00:07:46.189 --> 00:07:49.329 Two -sex islands. That is a pretty wild concept 00:07:49.329 --> 00:07:51.290 to think about. Try pointing Graphify at your 00:07:51.290 --> 00:07:53.269 messiest project this week just to see the map 00:07:53.269 --> 00:07:55.250 it generates. Thanks for joining us on this deep 00:07:55.250 --> 00:07:55.449 dive.