WEBVTT 00:00:00.000 --> 00:00:02.480 Imagine an AI that doesn't just, you know, pull 00:00:02.480 --> 00:00:06.379 up facts, but one that truly understands how 00:00:06.379 --> 00:00:08.640 those facts connect, how they influence each 00:00:08.640 --> 00:00:12.400 other. An AI that can reason, much like a brilliant 00:00:12.400 --> 00:00:15.199 research assistant, not just a super fast search 00:00:15.199 --> 00:00:17.620 engine. We're really talking about a fundamental 00:00:17.620 --> 00:00:21.379 shift in how AI interacts with information. It's 00:00:21.379 --> 00:00:23.800 like giving AI not just a memory, but a mind, 00:00:23.920 --> 00:00:26.239 you know, a way to connect things. This capability, 00:00:26.420 --> 00:00:29.460 it genuinely changes what's possible. Welcome 00:00:29.460 --> 00:00:31.699 to the Deep Dive. Today we're exploring a really 00:00:31.699 --> 00:00:34.880 cutting edge development in AI. It's this fusion 00:00:34.880 --> 00:00:38.579 of... agentic RG and knowledge graphs. Sounds 00:00:38.579 --> 00:00:41.500 fancy. It does. But our mission here is to unpack 00:00:41.500 --> 00:00:43.700 why the traditional ways AI retrieves information 00:00:43.700 --> 00:00:46.520 often, well, they hit a wall. Right. And then 00:00:46.520 --> 00:00:48.740 see how combining this sort of photographic memory 00:00:48.740 --> 00:00:51.460 with the ability to connect the dots creates 00:00:51.460 --> 00:00:54.880 something genuinely powerful. You'll get a feel 00:00:54.880 --> 00:00:57.240 for what enables this, how it's actually built, 00:00:57.299 --> 00:00:59.380 and what it could mean for how you interact with 00:00:59.380 --> 00:01:01.600 your own data. Yeah, and I think a lot of us 00:01:01.600 --> 00:01:04.280 have used AI to chat with documents, right? It 00:01:04.280 --> 00:01:05.920 feels pretty magical sometimes. It does. You 00:01:05.920 --> 00:01:08.319 ask a question, boom, it finds the answer right 00:01:08.319 --> 00:01:11.540 there in your files. But what happens when the 00:01:11.540 --> 00:01:14.620 questions get really complex, when you need to 00:01:14.620 --> 00:01:18.180 understand relationships, maybe compare strategies 00:01:18.180 --> 00:01:20.739 across different sources? That's where it gets 00:01:20.739 --> 00:01:23.730 tricky. Right. So for anyone who's played with 00:01:23.730 --> 00:01:25.670 traditional RRAG, that's retrieval augmented 00:01:25.670 --> 00:01:27.950 generation, it's a good starting point. It's 00:01:27.950 --> 00:01:30.510 meant to ground a language model in your specific 00:01:30.510 --> 00:01:34.430 documents. But what are its limits? When do questions 00:01:34.430 --> 00:01:37.030 need more than just a simple lookup? Well, think 00:01:37.030 --> 00:01:40.209 of it like this. Traditional RRAG, it's great 00:01:40.209 --> 00:01:42.510 for direct answers, finding that specific sentence. 00:01:42.689 --> 00:01:45.420 But it's fundamentally inflexible. The process 00:01:45.420 --> 00:01:47.560 usually involves breaking documents into these 00:01:47.560 --> 00:01:50.859 little chunks. Then you create mathematical embeddings, 00:01:50.939 --> 00:01:53.099 sort of numerical fingerprints for each chunk, 00:01:53.299 --> 00:01:55.859 and store them in what's called a vector database. 00:01:56.140 --> 00:01:59.299 This is for quick similarity searches. So when 00:01:59.299 --> 00:02:02.019 you ask a question, the AI just finds the most 00:02:02.019 --> 00:02:04.579 similar chunks and stuffs them into the prompt 00:02:04.579 --> 00:02:07.159 for the language model. So it gets the context, 00:02:07.459 --> 00:02:09.840 the raw text, but it can't really reason about 00:02:09.840 --> 00:02:11.360 it. It doesn't understand the meaning behind 00:02:11.360 --> 00:02:15.289 it. Precisely. The AI is constrained. It can't 00:02:15.289 --> 00:02:17.509 say, look at the initial search results and think, 00:02:17.610 --> 00:02:19.930 hmm, these aren't quite right. Let me try a different 00:02:19.930 --> 00:02:23.949 angle. It's not built to explore complex relationships 00:02:23.949 --> 00:02:27.050 or do multi -step reasoning across different 00:02:27.050 --> 00:02:29.650 facts it finds. So essentially, it ends up being 00:02:29.650 --> 00:02:31.919 a very sophisticated search engine. But still 00:02:31.919 --> 00:02:34.240 just a search engine. Yeah, exactly. And one 00:02:34.240 --> 00:02:35.840 that's kind of blind to the deeper connections. 00:02:35.919 --> 00:02:38.400 It's like someone reading random paragraphs from 00:02:38.400 --> 00:02:40.740 a dozen books, but never grasping the overall 00:02:40.740 --> 00:02:43.439 story. Okay, so what specific kinds of questions 00:02:43.439 --> 00:02:47.340 really break these traditional ARG agents? Where 00:02:47.340 --> 00:02:49.500 do they really stumble? Questions about relationships, 00:02:49.780 --> 00:02:52.759 comparisons, or maybe cause and effect. Those 00:02:52.759 --> 00:02:55.319 often cause major trouble. Got it. Questions 00:02:55.319 --> 00:02:57.460 needing more than just finding similar words. 00:02:57.719 --> 00:03:00.139 Yeah, things that require synthesis. Right. Okay. 00:03:00.240 --> 00:03:01.719 So here's where it gets really interesting then. 00:03:02.080 --> 00:03:05.740 This new approach, agentic R, combined with knowledge 00:03:05.740 --> 00:03:08.860 graphs, you said it gives AI not just that memory, 00:03:08.919 --> 00:03:11.240 but also this ability to connect the dots. That 00:03:11.240 --> 00:03:13.680 really does sound like a big leap. It truly is. 00:03:14.120 --> 00:03:16.979 Think of it like this. If traditional R is a 00:03:16.979 --> 00:03:19.659 really, really good filing cabinet, excellent 00:03:19.659 --> 00:03:23.080 at pulling out specific files quickly, then this 00:03:23.080 --> 00:03:26.120 new system... is more like your brilliant research 00:03:26.120 --> 00:03:28.919 assistant. This assistant hasn't just read every 00:03:28.919 --> 00:03:30.860 document you gave them. They've actually understood 00:03:30.860 --> 00:03:33.560 how everything within those documents relates 00:03:33.560 --> 00:03:35.800 to everything else. They've built this comprehensive 00:03:35.800 --> 00:03:38.919 mental map. So the AI itself, the agent, it has 00:03:38.919 --> 00:03:41.060 autonomy. It gets to decide the best strategy 00:03:41.060 --> 00:03:42.919 to answer my question instead of just following 00:03:42.919 --> 00:03:46.300 one rigid search then answer path. Exactly. That's 00:03:46.300 --> 00:03:48.580 the core idea. For a simple fact, like what year 00:03:48.580 --> 00:03:51.039 was this company founded? It might just do a 00:03:51.039 --> 00:03:53.680 quick vector search. Fast and easy. If you ask 00:03:53.680 --> 00:03:57.360 about relationships, like how does this new regulation 00:03:57.360 --> 00:04:00.919 impact our main competitor, It can traverse the 00:04:00.919 --> 00:04:02.759 knowledge graph, mapping out those connections. 00:04:02.979 --> 00:04:05.960 And for really complex queries, it might intelligently 00:04:05.960 --> 00:04:09.460 combine both methods. Two -sec silence, you know, 00:04:09.460 --> 00:04:11.780 a real -world example. Instead of just finding 00:04:11.780 --> 00:04:14.479 separate documents mentioning Microsoft and OpenAI, 00:04:14.740 --> 00:04:17.000 this agent, because it has that knowledge graph, 00:04:17.120 --> 00:04:19.740 it understands their multibillion -dollar partnership. 00:04:19.920 --> 00:04:22.899 It knows how that connection impacts their strategies, 00:04:22.980 --> 00:04:25.220 their products, everything. It's connected knowledge. 00:04:25.769 --> 00:04:27.750 not just isolated facts. And that's what you 00:04:27.750 --> 00:04:30.610 mean by agentic argue, the agent choosing the 00:04:30.610 --> 00:04:33.610 strategy. Yep. An AI system where the agent itself 00:04:33.610 --> 00:04:36.189 figures out the best way to find and use information 00:04:36.189 --> 00:04:39.629 to answer your query. So how does this deeper 00:04:39.629 --> 00:04:42.970 level of understanding fundamentally change what 00:04:42.970 --> 00:04:45.430 AI can do for us? What kinds of problems can 00:04:45.430 --> 00:04:48.009 it tackle now? Well, it allows the AI to actually 00:04:48.009 --> 00:04:50.629 reason and provide much more nuanced answers 00:04:50.629 --> 00:04:53.670 going way beyond simple fact retrieval. It starts 00:04:53.670 --> 00:04:55.829 to sound like actual understanding. It's getting 00:04:55.829 --> 00:04:59.149 closer. This capability, it sounds pretty advanced. 00:04:59.449 --> 00:05:01.870 What are the actual tools making this possible? 00:05:02.029 --> 00:05:04.910 Is this all super proprietary, complex stuff? 00:05:05.129 --> 00:05:07.790 Surprisingly, no. A lot of it is built on open 00:05:07.790 --> 00:05:09.290 source components. We're talking about things 00:05:09.290 --> 00:05:12.029 like Pydantic AI. That's a Python library that 00:05:12.029 --> 00:05:14.389 helps build structured, reliable outputs from 00:05:14.389 --> 00:05:16.709 the AI agent. So it's not just rambling. It gives 00:05:16.709 --> 00:05:18.529 you organized answers. Okay, structure is good. 00:05:18.730 --> 00:05:21.050 Yeah. And for the knowledge graph itself, there's 00:05:21.050 --> 00:05:24.139 a tool called Graffiti. to help define the structure, 00:05:24.259 --> 00:05:28.240 the schema. And then Neo4j is pretty much the 00:05:28.240 --> 00:05:31.139 leader for graph databases where all those intricate 00:05:31.139 --> 00:05:33.339 connections actually live and get queried. Right, 00:05:33.439 --> 00:05:35.639 Neo4j. And what about the vector database piece 00:05:35.639 --> 00:05:38.459 for that initial similarity search? Ah, yeah, 00:05:38.579 --> 00:05:40.759 that's a pretty clever choice in this setup. 00:05:40.879 --> 00:05:44.480 It uses PostgreSQL, the familiar, robust SQL 00:05:44.480 --> 00:05:47.839 database, but with an extension called pgVector. 00:05:48.250 --> 00:05:50.550 So your standard relational database suddenly 00:05:50.550 --> 00:05:52.750 gains high -performance vector search capabilities. 00:05:52.970 --> 00:05:55.769 Whoa, wait. Postgresible with PGVector? So it's 00:05:55.769 --> 00:05:58.050 like a database and a vector search engine rolled 00:05:58.050 --> 00:06:00.550 into one, a two -in -one superpower? Yeah, kind 00:06:00.550 --> 00:06:02.209 of like having a sports car that also happens 00:06:02.209 --> 00:06:04.290 to be a submarine. Yeah. It's really powerful. 00:06:04.310 --> 00:06:07.129 You get this solid, reliable SQL foundation, 00:06:07.189 --> 00:06:09.930 but can also handle the vector stuff extremely 00:06:09.930 --> 00:06:12.689 well for many use cases. That sounds incredibly 00:06:12.689 --> 00:06:15.699 versatile. But I wonder, does putting vector 00:06:15.699 --> 00:06:18.560 search inside PostgreSQL have any trade -offs 00:06:18.560 --> 00:06:21.939 compared to using, say, a totally separate specialized 00:06:21.939 --> 00:06:25.600 vector database? That's a fair question. You 00:06:25.600 --> 00:06:27.600 definitely gain a lot of convenience and often 00:06:27.600 --> 00:06:29.720 simplify your whole setup. You have one less 00:06:29.720 --> 00:06:32.980 system to manage. But, you know, for scenarios 00:06:32.980 --> 00:06:35.459 demanding absolutely massive scale, maybe billions 00:06:35.459 --> 00:06:39.449 of vectors or ultra -low latency search. a dedicated 00:06:39.449 --> 00:06:41.850 specialized vector database might still have 00:06:41.850 --> 00:06:44.509 an edge. Okay. For most applications, though, 00:06:44.569 --> 00:06:46.550 this Swiss army knife approach with Postgres 00:06:46.550 --> 00:06:49.389 and PG vector is incredibly effective and flexible. 00:06:49.529 --> 00:06:52.250 Plus, the whole thing uses fast API for the API 00:06:52.250 --> 00:06:55.209 layer, which is modern and fast. Right. And critically 00:06:55.209 --> 00:06:57.870 important, I think, you have amazing flexibility 00:06:57.870 --> 00:07:00.529 with the large language models of brains, right? 00:07:01.000 --> 00:07:04.459 You can plug in OpenAI's models or run open source 00:07:04.459 --> 00:07:07.620 models locally using Alama or use Google Gemini 00:07:07.620 --> 00:07:10.139 models. You're not locked into one vendor. No, 00:07:10.279 --> 00:07:12.959 vendor lock -in is huge. So what's the key benefit, 00:07:13.160 --> 00:07:15.699 the sort of magic, when you combine these specific 00:07:15.699 --> 00:07:19.160 tools, Pydantic AI, Graffiti, Neo4j, Postgres 00:07:19.160 --> 00:07:21.839 Vector, FastAPI? It's that seamless integration 00:07:21.839 --> 00:07:24.319 that allows you to build these really powerful 00:07:24.319 --> 00:07:28.540 yet customizable AI systems that can both search 00:07:28.540 --> 00:07:31.540 and reason. Powerful and customizable. I like 00:07:31.540 --> 00:07:34.139 the sound of that. Sponsoreed mid -roll. Okay, 00:07:34.180 --> 00:07:35.920 so how does someone actually build this kind 00:07:35.920 --> 00:07:37.399 of assistant? It sounds like a lot of moving 00:07:37.399 --> 00:07:40.100 parts, but the article suggests it's pretty approachable 00:07:40.100 --> 00:07:42.379 if you're willing to dive in. Yeah, the article 00:07:42.379 --> 00:07:44.000 lays it out step by step. You basically get the 00:07:44.000 --> 00:07:46.100 code, set up your Python environment. You'll 00:07:46.100 --> 00:07:48.480 need PostgreSQL. And they suggest maybe using 00:07:48.480 --> 00:07:50.860 a managed service like Neon, which handles the 00:07:50.860 --> 00:07:53.079 vector stuff easily. Okay, Neon for Postgres. 00:07:53.319 --> 00:07:57.040 And then your Neo4j instance for the graph database, 00:07:57.300 --> 00:08:01.180 Beat. But the really crucial part, the magic 00:08:01.180 --> 00:08:03.579 moment, is loading your knowledge base. Right, 00:08:03.639 --> 00:08:05.819 your documents. Exactly. You just put your documents, 00:08:06.000 --> 00:08:08.620 PDFs, text files, whatever, into a specific folder. 00:08:08.759 --> 00:08:11.000 Then you run an ingestion script, and the AI 00:08:11.000 --> 00:08:13.279 gets to work. What's it doing, though? It splits 00:08:13.279 --> 00:08:15.800 the text into meaningful chunks, not just random 00:08:15.800 --> 00:08:19.160 bits. It generates those vector embeddings for 00:08:19.160 --> 00:08:21.540 the similarity search in Postgresql. And this 00:08:21.540 --> 00:08:24.680 is key. It uses a large language model to read 00:08:24.680 --> 00:08:27.500 through the text and identify the important entities 00:08:27.500 --> 00:08:30.540 like people, companies, concepts, and the relationships 00:08:30.540 --> 00:08:34.000 between them. Ah, so the LLM itself helps build 00:08:34.000 --> 00:08:36.860 the knowledge graph. Precisely. It extracts that 00:08:36.860 --> 00:08:39.019 structure, that understanding, and populates 00:08:39.019 --> 00:08:41.340 the Neo4j graph database. That's where the real 00:08:41.340 --> 00:08:43.279 understanding of your data gets built. Okay, 00:08:43.340 --> 00:08:45.720 so once that's built and the graph is populated... 00:08:46.139 --> 00:08:48.559 How do we test it? What kinds of questions really 00:08:48.559 --> 00:08:51.639 show off its reasoning beyond just finding facts? 00:08:52.080 --> 00:08:54.120 This is the fun part. You can actually see its 00:08:54.120 --> 00:08:56.460 logic. Ask it something simple like, what are 00:08:56.460 --> 00:08:59.120 Google's main AI initiatives? The agent will 00:08:59.120 --> 00:09:01.980 probably just use a quick vector search. Finds 00:09:01.980 --> 00:09:04.419 the relevant passages. Simple. Right. But then 00:09:04.419 --> 00:09:06.960 ask something like, how are Amazon and Anthropic 00:09:06.960 --> 00:09:09.980 connected? Now, it gets interesting. The agent 00:09:09.980 --> 00:09:11.659 should realize this is about a relationship. 00:09:12.669 --> 00:09:15.029 So it uses the graph search. It navigates the 00:09:15.029 --> 00:09:18.269 connections in Neo4j and finds their big investment 00:09:18.269 --> 00:09:22.070 relationship, maybe Anthropic's use of AWS infrastructure. 00:09:22.450 --> 00:09:24.809 It pieces together the connection. That's cool. 00:09:25.009 --> 00:09:27.070 And then for the really complex stuff, multi 00:09:27.070 --> 00:09:30.450 -step reasoning. Ask it something like, compare 00:09:30.450 --> 00:09:33.450 Microsoft's AI strategy with OpenAI's approach. 00:09:33.830 --> 00:09:35.950 Okay, that's not a simple lookup. Not at all. 00:09:36.009 --> 00:09:38.889 Beep, beep. Here, the agent has to be smart. 00:09:39.009 --> 00:09:41.429 It needs to combine vector search to pull specific 00:09:41.429 --> 00:09:44.429 facts about each company's strategy and graph 00:09:44.429 --> 00:09:46.309 search to understand their deep partnership, 00:09:46.590 --> 00:09:49.129 how one influences the other. The result should 00:09:49.129 --> 00:09:51.870 be a really nuanced, comprehensive answer that 00:09:51.870 --> 00:09:54.049 considers both the individual parts and their 00:09:54.049 --> 00:09:56.230 relationship. It's genuinely impressive when 00:09:56.230 --> 00:09:59.190 you see it work. Wow. Imagine scaling that kind 00:09:59.190 --> 00:10:01.669 of analysis. A billion queries like that. The 00:10:01.669 --> 00:10:04.389 depth of understanding you could unlock. It's 00:10:04.389 --> 00:10:07.179 incredible. Really remarkable potential there. 00:10:07.299 --> 00:10:10.279 So thinking about getting it running, beyond 00:10:10.279 --> 00:10:12.799 just installing the software and setting environment 00:10:12.799 --> 00:10:15.799 variables, what's often the trickiest part in 00:10:15.799 --> 00:10:18.379 getting the quality right, making sure the understanding 00:10:18.379 --> 00:10:21.340 is actually good. Yeah, that's key. Setting up 00:10:21.340 --> 00:10:23.259 the databases and making sure all the connections 00:10:23.259 --> 00:10:25.840 and like. API keys are right, can definitely 00:10:25.840 --> 00:10:28.179 be tricky for first timers. But getting the quality 00:10:28.179 --> 00:10:31.340 often comes down to the data feeding in and how 00:10:31.340 --> 00:10:33.460 the agent reasons. Right, the inputs matter. 00:10:33.740 --> 00:10:36.740 Hugely. Okay, so once a system is built, how 00:10:36.740 --> 00:10:39.720 locked in are you? How customizable is it? Can 00:10:39.720 --> 00:10:42.139 you really tailor it? Oh, absolutely. That's 00:10:42.139 --> 00:10:44.500 the beauty of it being mostly open source. You 00:10:44.500 --> 00:10:46.519 can totally customize the types of entities and 00:10:46.519 --> 00:10:48.399 relationships it looks for. Maybe you need it 00:10:48.399 --> 00:10:50.980 to understand legal clauses or specific scientific 00:10:50.980 --> 00:10:53.940 concepts. You could integrate multimodal knowledge. 00:10:54.730 --> 00:10:57.649 Imagine feeding it images or audio files alongside 00:10:57.649 --> 00:11:00.809 text. You can set up real -time updates so as 00:11:00.809 --> 00:11:02.950 new documents come in, the knowledge graph grows. 00:11:03.210 --> 00:11:06.470 And you can definitely optimize performance for 00:11:06.470 --> 00:11:09.009 really massive data sets. It's designed to be 00:11:09.009 --> 00:11:11.429 flexible. That's really powerful. Now, what's 00:11:11.429 --> 00:11:13.230 equally fascinating, you mentioned earlier, is 00:11:13.230 --> 00:11:16.190 how the system itself was actually built. The 00:11:16.190 --> 00:11:18.190 author used something called context engineering 00:11:18.190 --> 00:11:21.690 with an advanced AI like Claude. So the AI wasn't 00:11:21.690 --> 00:11:23.789 just like a co -pilot helping out. It was more 00:11:23.789 --> 00:11:25.909 fundamental. Way more than a co -pilot, really. 00:11:26.029 --> 00:11:29.299 The description makes it sound like... Imagine 00:11:29.299 --> 00:11:32.279 the AI was the incredibly skilled camera operator, 00:11:32.460 --> 00:11:35.340 the lighting expert, the sound engineer, and 00:11:35.340 --> 00:11:38.019 the editor. Wow. While the human was primarily 00:11:38.019 --> 00:11:40.279 the film director, providing the vision and the 00:11:40.279 --> 00:11:43.620 overall plan. Apparently, over 90 % of the system's 00:11:43.620 --> 00:11:46.039 code was generated by the AI in just 35 minutes. 00:11:46.139 --> 00:11:50.179 35 minutes. Yeah. By feeding it a really detailed 00:11:50.179 --> 00:11:53.379 project specification, the requirements, and 00:11:53.379 --> 00:11:55.980 crucially, access to the documentation for all 00:11:55.980 --> 00:11:58.440 those tools we mentioned. vulnerable admission 00:11:58.440 --> 00:12:01.519 I mean I still wrestle with prompt drift myself 00:12:01.519 --> 00:12:04.279 sometimes getting the AI to consistently do what 00:12:04.279 --> 00:12:07.019 I want so seeing this level of complex system 00:12:07.019 --> 00:12:10.100 generation through AI assistance is genuinely 00:12:10.100 --> 00:12:12.320 mind -blowing and that's context engineering 00:12:12.320 --> 00:12:14.759 giving the AI all that background material essentially 00:12:14.759 --> 00:12:18.440 yeah it's about guiding the AI assistant by providing 00:12:18.440 --> 00:12:21.600 comprehensive documentation clear examples detailed 00:12:21.600 --> 00:12:24.850 plans basically giving it everything it needs 00:12:24.850 --> 00:12:27.070 to understand the task deeply and generate the 00:12:27.070 --> 00:12:29.710 right code or system components. So what's the 00:12:29.710 --> 00:12:32.950 big takeaway then from seeing AI build such a 00:12:32.950 --> 00:12:35.389 sophisticated system so rapidly? What does that 00:12:35.389 --> 00:12:37.789 imply more broadly? It just shows how AI -assisted 00:12:37.789 --> 00:12:39.990 development can drastically speed up the creation 00:12:39.990 --> 00:12:42.490 of complex systems. It's a game changer for building 00:12:42.490 --> 00:12:44.470 things quickly. A definite force multiplier. 00:12:44.809 --> 00:12:46.789 Okay, so for listeners who are thinking, hey, 00:12:46.809 --> 00:12:48.710 I want to try building this, are there any common 00:12:48.710 --> 00:12:51.750 pitfalls, any gotchas they should watch out for 00:12:51.750 --> 00:12:53.759 that might trip them up? Definitely a few things 00:12:53.759 --> 00:12:56.899 to keep in mind. First, during your initial testing 00:12:56.899 --> 00:12:59.620 phase, creating that knowledge graph from scratch 00:12:59.620 --> 00:13:02.360 can sometimes take a really long time, especially 00:13:02.360 --> 00:13:05.240 with lots of documents. Right. The good news 00:13:05.240 --> 00:13:07.639 is you can actually skip that step temporarily 00:13:07.639 --> 00:13:10.659 for faster iteration. There's usually a flag 00:13:10.659 --> 00:13:12.480 you can use when running the ingestion script, 00:13:12.720 --> 00:13:15.220 something like no knowledge graph, just to get 00:13:15.220 --> 00:13:18.019 the vector search part working first. Okay. Handy 00:13:18.019 --> 00:13:20.519 tip for testing. What else? Another common one. 00:13:20.779 --> 00:13:23.080 Sometimes you find the agent gets stuck in a 00:13:23.080 --> 00:13:26.600 rut. It always uses vector search or always tries 00:13:26.600 --> 00:13:29.419 the graph, even when the question clearly suits 00:13:29.419 --> 00:13:32.100 the other method. Ah, it's not making the right 00:13:32.100 --> 00:13:34.980 decision. Exactly. If that happens, you need 00:13:34.980 --> 00:13:37.480 to go back and refine the system prompt you give 00:13:37.480 --> 00:13:39.940 the agent. You have to be super clear about the 00:13:39.940 --> 00:13:42.940 criteria it should use to decide which tool vector 00:13:42.940 --> 00:13:45.000 search or graph search is appropriate for different 00:13:45.000 --> 00:13:47.659 kinds of questions. So tweaking the agent's instructions. 00:13:50.649 --> 00:13:53.070 off irrelevant like it's not finding the right 00:13:53.070 --> 00:13:56.190 chunks yeah that can happen first thing is maybe 00:13:56.190 --> 00:13:58.509 experiment with different embedding models some 00:13:58.509 --> 00:14:00.409 are better than others for certain types of text 00:14:00.409 --> 00:14:04.210 okay but crucially pay attention to how you're 00:14:04.210 --> 00:14:06.330 chunking the documents this idea of semantic 00:14:06.330 --> 00:14:08.809 chunking is really important semantic chunking 00:14:08.809 --> 00:14:10.570 yeah instead of just chopping documents into 00:14:10.570 --> 00:14:15.639 rigid say 500 word blocks right You split them 00:14:15.639 --> 00:14:17.919 based on logical breaks in the content, paragraph 00:14:17.919 --> 00:14:21.120 sections, topic shifts, meaningful segments. 00:14:21.259 --> 00:14:23.159 That usually gives the vector search much better 00:14:23.159 --> 00:14:25.639 context to work with. Semantic chunking basically 00:14:25.639 --> 00:14:28.919 is just dividing text into meaningful, logically 00:14:28.919 --> 00:14:31.960 coherent parts, not just fixed size blocks. It 00:14:31.960 --> 00:14:34.059 makes a big difference. Makes sense. Better chunks, 00:14:34.220 --> 00:14:37.299 better search. And what about if you just can't 00:14:37.299 --> 00:14:39.580 connect to Neo4j, the graph database? Oh yeah, 00:14:39.600 --> 00:14:41.659 that's usually simpler. Double check your credentials, 00:14:41.879 --> 00:14:44.759 username, password, database URL, and just make 00:14:44.759 --> 00:14:46.679 sure the Neo4j service is actually running on 00:14:46.679 --> 00:14:49.059 your machine or wherever it's hosted. It's often 00:14:49.059 --> 00:14:51.039 just a configuration type or the service isn't 00:14:51.039 --> 00:14:54.679 started. Standard troubleshooting then. So beyond 00:14:54.679 --> 00:14:57.539 these specific fixes, what's a good general mindset 00:14:57.539 --> 00:14:59.620 to have when you're troubleshooting these complex 00:14:59.620 --> 00:15:04.039 AI systems? Patience. iteration it often involves 00:15:04.039 --> 00:15:06.500 carefully tweaking the AI's internal prompts 00:15:06.500 --> 00:15:09.220 trying different chunking strategies maybe adjusting 00:15:09.220 --> 00:15:11.820 parameters and seeing how it affects the results 00:15:11.820 --> 00:15:15.659 it's rarely a one -shot fix careful refinement 00:15:15.659 --> 00:15:19.460 is key iteration and refinement got it so let's 00:15:19.460 --> 00:15:21.159 pull back what does this all mean for us what's 00:15:21.159 --> 00:15:23.799 the really big idea here it feels like more than 00:15:23.799 --> 00:15:27.149 just a new technique I think it is. The big idea 00:15:27.149 --> 00:15:29.149 is that the future of AI, especially when it 00:15:29.149 --> 00:15:31.169 comes to understanding information, isn't just 00:15:31.169 --> 00:15:33.490 about making search faster or training even bigger 00:15:33.490 --> 00:15:36.149 language models. It's really about building systems 00:15:36.149 --> 00:15:38.450 that can genuinely reason. We're moving away 00:15:38.450 --> 00:15:40.450 from thinking about static databases that just 00:15:40.450 --> 00:15:43.610 store isolated facts towards building dynamic, 00:15:43.649 --> 00:15:46.809 interconnected knowledge networks where the relationships 00:15:46.809 --> 00:15:49.330 are just as important as the facts themselves. 00:15:50.039 --> 00:15:52.139 You know, traditional reg gives you a really 00:15:52.139 --> 00:15:54.940 good search engine. Agetic reg, combined with 00:15:54.940 --> 00:15:57.039 knowledge graphs, starts to give you something 00:15:57.039 --> 00:15:59.340 closer to a brilliant research assistant. From 00:15:59.340 --> 00:16:01.500 search engine to research assistant, that's a 00:16:01.500 --> 00:16:04.620 powerful shift. It's a fundamental change in 00:16:04.620 --> 00:16:07.559 how we can expect AI to help us make sense of 00:16:07.559 --> 00:16:10.500 complex information. So this deep dive really 00:16:10.500 --> 00:16:13.159 reveals a fundamental shift, doesn't it? Moving 00:16:13.159 --> 00:16:15.440 from just simple search towards an intelligent 00:16:15.440 --> 00:16:18.600 understanding that actually grasps context and 00:16:18.600 --> 00:16:21.899 crucially relationships. Absolutely. And the 00:16:21.899 --> 00:16:24.019 tools are becoming accessible enough to actually 00:16:24.019 --> 00:16:26.220 build this now. So maybe some homework for our 00:16:26.220 --> 00:16:28.100 listeners, if you're feeling ambitious, that 00:16:28.100 --> 00:16:30.899 is. Yeah. Get your hands dirty. Try setting this 00:16:30.899 --> 00:16:32.960 system up. The article provides the guide and 00:16:32.960 --> 00:16:35.259 the code. Load it with your own documents. Maybe 00:16:35.259 --> 00:16:37.600 it's your company's internal notes. Maybe research 00:16:37.600 --> 00:16:40.100 papers you need to synthesize. Project documentation. 00:16:40.700 --> 00:16:43.360 Whatever complex information you need to master. 00:16:43.519 --> 00:16:46.200 And then test it. Exactly. Ask it different kinds 00:16:46.200 --> 00:16:49.019 of questions, simple facts, relationship questions. 00:16:49.120 --> 00:16:52.379 Those complex comparison queries push its boundaries. 00:16:52.620 --> 00:16:54.460 See what it can do with your data. Because this 00:16:54.460 --> 00:16:56.500 is really about more than just playing with code, 00:16:56.600 --> 00:16:59.539 isn't it? It's about starting to grasp how AI 00:16:59.539 --> 00:17:03.120 can navigate that incredibly complex web of relationships 00:17:03.120 --> 00:17:05.259 within your information, uncovering insights 00:17:05.259 --> 00:17:07.960 that were maybe hidden before. It's about starting 00:17:07.960 --> 00:17:10.920 to build that future of AI for yourself, tailored 00:17:10.920 --> 00:17:13.319 to what you need to understand. So the final 00:17:13.319 --> 00:17:15.279 thought to leave... everyone with, what kind 00:17:15.279 --> 00:17:17.740 of complex relationship -based questions could 00:17:17.740 --> 00:17:20.859 you finally get answers to if your AI truly understood 00:17:20.859 --> 00:17:22.799 your data? Out to your own music.