WEBVTT 00:05.000 --> 00:11.280 Large language models are extremely powerful, but they have a fundamental limitation that often gets 00:11.280 --> 00:12.000 overlooked. 00:12.560 --> 00:20.080 They generate responses solely from their training data and learn statistical patterns, as shown on 00:20.080 --> 00:21.360 page two of the deck. 00:21.840 --> 00:27.720 They do not query live systems, verify facts, or check whether information is current. 00:28.680 --> 00:33.480 This means that from the moment training ends, the model's knowledge begins to age. 00:33.960 --> 00:40.560 It may confidently reference outdated policies, discontinued products, or obsolete procedures. 00:41.120 --> 00:46.720 Worse, when the model encounters gaps in its knowledge, it does not say I don't know. 00:47.000 --> 00:51.000 Instead, it generates the most plausible sounding answer it can. 00:51.560 --> 00:59.000 This leads to three serious problems hallucinations, outdated information, and confidently wrong answers. 00:59.760 --> 01:05.230 Because the language is fluent, Users often trust responses that are factually incorrect. 01:05.910 --> 01:11.670 There is also no inherent source verification claims cannot be traced or audited. 01:11.830 --> 01:19.550 The key insight is critical for engineers to internalize Llms are exceptional language models, but 01:19.550 --> 01:23.710 they are not knowledge databases without external grounding. 01:23.870 --> 01:27.950 They are unreliable as information systems in production environments. 01:28.150 --> 01:28.910 Retrieval. 01:28.910 --> 01:35.510 Augmented generation, or Rag, is an architectural pattern designed to solve the core weaknesses of 01:35.510 --> 01:36.990 standalone llms. 01:37.750 --> 01:44.430 As illustrated on page three of the Dec rag fundamentally changes how models access and use information. 01:45.030 --> 01:51.630 Instead of relying entirely on parametric memory, what the model learned during training, Rag introduces 01:51.670 --> 01:53.910 a retrieval step before generation. 01:54.470 --> 02:00.750 First, the system retrieves relevant documents from a knowledge base, vector database or document 02:00.750 --> 02:01.270 store. 02:01.790 --> 02:06.350 Next, those documents are injected directly into the prompt as context. 02:06.910 --> 02:11.710 Finally, the LLM generates a response grounded in that retrieved information. 02:12.310 --> 02:18.630 This architecture allows the model to combine its natural language capabilities with external verifiable 02:18.630 --> 02:19.270 knowledge. 02:19.910 --> 02:26.070 The result is not just more accurate answers, but answers that can be traced back to specific sources. 02:26.790 --> 02:31.550 The most important takeaway is that Rag is not a prompt trick. 02:32.110 --> 02:34.470 It is a system level design pattern. 02:34.990 --> 02:43.030 By separating language capability from knowledge storage, Rag enables Llms to operate as reliable information 02:43.030 --> 02:46.950 synthesis engines rather than guess based text generators. 02:47.670 --> 02:54.910 Hallucinations occur because llms are trained to predict the next most likely token not to verify truth. 02:55.470 --> 03:00.510 When the model lacks information, it fills the gap with statistically plausible text. 03:00.550 --> 03:03.430 Page four of the deck explains this clearly. 03:03.670 --> 03:07.860 Hallucinations are an inevitable outcome of the training objective itself. 03:08.340 --> 03:13.260 Rag addresses this problem by changing the conditions under which generation happens. 03:13.740 --> 03:15.740 Instead of forcing the model to guess. 03:15.940 --> 03:20.420 Rag supplies relevant factual context before generation begins. 03:20.700 --> 03:27.060 This retrieved context fills knowledge gaps and narrows the model's output space by anchoring responses 03:27.060 --> 03:28.260 to real documents. 03:28.420 --> 03:32.340 Rag reduces the model's tendency to fabricate information. 03:32.860 --> 03:37.500 Claims can be traced back to source material, making verification possible. 03:38.180 --> 03:40.500 The model no longer operates in a vacuum. 03:40.820 --> 03:43.100 It is guided by concrete evidence. 03:43.180 --> 03:48.540 This represents a fundamental engineering shift from guessing to grounding. 03:49.020 --> 03:56.180 With Rag llms transition from creative language generators into controlled information synthesis systems. 03:56.740 --> 04:03.340 While hallucinations cannot be eliminated entirely, Rag dramatically reduces their frequency and impact, 04:03.700 --> 04:07.380 making production use far safer and more reliable. 04:07.580 --> 04:12.980 One of the most serious limitations of standalone LMS is static knowledge. 04:13.420 --> 04:19.020 As highlighted on page five of the Dec training, data is frozen at a specific point in time. 04:19.820 --> 04:27.260 A model trained in 2023 has no awareness of events, policies or product changes that occur afterward. 04:27.460 --> 04:33.300 Traditional solutions attempt to solve this with fine tuning or retraining, but those approaches are 04:33.300 --> 04:36.900 expensive, slow, and operationally complex. 04:37.620 --> 04:44.700 Rag provides a fundamentally better solution by decoupling knowledge from model parameters with Rag 04:44.900 --> 04:52.300 knowledge lives in external systems, databases, document stores, APIs that can be updated independently, 04:52.780 --> 04:58.820 new documents can be added, outdated ones removed, and corrections applied without touching the underlying 04:58.860 --> 04:59.420 model. 05:00.060 --> 05:06.100 This enables real time knowledge access and continuous improvement systems can integrate. 05:06.100 --> 05:13.170 Live data sources reflect policy updates immediately and remain accurate without costly retraining cycles. 05:13.330 --> 05:16.530 For enterprises, this capability is critical. 05:16.970 --> 05:19.250 Business knowledge changes constantly. 05:19.650 --> 05:26.890 Rag ensures AI systems stay current, relevant and aligned with reality rather than locked to a historical 05:26.890 --> 05:27.690 snapshot. 05:28.410 --> 05:34.690 Enterprise environments introduce challenges that make Rag not just useful, but essential. 05:35.410 --> 05:37.410 As shown on page six of the Dec. 05:37.770 --> 05:42.290 Enterprise data is private, proprietary, and domain specific. 05:42.850 --> 05:45.410 It cannot be included in public model training. 05:45.410 --> 05:52.930 Data organizations rely on specialized terminology, internal processes, and institutional knowledge 05:52.930 --> 05:54.330 that exist nowhere else. 05:54.970 --> 05:58.170 At the same time, this information changes frequently. 05:58.410 --> 06:03.010 Policies, procedures, pricing, and regulations may be updated daily. 06:03.730 --> 06:07.570 Standalone llms cannot access this internal knowledge. 06:08.200 --> 06:14.600 Even worse, relying on them without grounding risks, misinformation, compliance violations and loss 06:14.600 --> 06:15.320 of trust. 06:16.000 --> 06:22.480 Wragg bridges this gap by allowing AI systems to retrieve and reason over private enterprise data, 06:22.480 --> 06:25.160 while maintaining security and access control. 06:25.800 --> 06:31.400 Knowledge remains within organizational boundaries, and the model only sees what it is allowed to see 06:31.440 --> 06:34.120 at query time for enterprises. 06:34.400 --> 06:41.560 Wragg is the difference between experimental AI and deployable, trustworthy systems that meet real 06:41.600 --> 06:43.040 business requirements. 06:43.080 --> 06:50.400 Wragg enables a wide range of high impact enterprise AI use cases, as outlined on page seven of the 06:50.400 --> 06:50.840 deck. 06:51.520 --> 06:57.840 One of the most common is internal knowledge assistance, which help employees find information across 06:57.840 --> 07:02.280 wikis, documentation, and policy repositories instantly. 07:03.080 --> 07:06.080 Customer support bots are another major application. 07:06.480 --> 07:12.040 By retrieving from product documentation, troubleshooting guides and historical support tickets. 07:12.480 --> 07:18.520 Rag powered bots provide accurate, context aware responses instead of generic answers. 07:19.160 --> 07:23.480 Compliance and policy search is especially important in regulated industries. 07:23.960 --> 07:31.080 Rag allows employees to query regulatory requirements and internal policies while maintaining auditability 07:31.080 --> 07:32.200 and traceability. 07:32.960 --> 07:39.920 Technical documentation Q&A helps engineers navigate complex APIs and architectures efficiently. 07:40.160 --> 07:46.880 Analytics and reporting assistants use Rag to retrieve relevant metrics and generate natural language 07:46.880 --> 07:50.160 explanations across all these use cases. 07:50.360 --> 07:56.360 One requirement is universal answers must be accurate, auditable, and traceable. 07:56.920 --> 08:01.280 Rag provides the architecture needed to meet these enterprise standards. 08:01.400 --> 08:06.840 Rag is often compared with fine tuning, but they solve very different problems. 08:07.280 --> 08:13.470 As summarized on page eight of the Dec, fine tuning embeds knowledge directly into model weights. 08:13.910 --> 08:19.350 This makes updates costly and risky, especially when knowledge changes frequently. 08:19.830 --> 08:20.510 Rag. 08:20.550 --> 08:23.350 By contrast, keeps knowledge external. 08:23.790 --> 08:29.590 It is easier to update, lower risk, and far more suitable for enterprise environments. 08:30.110 --> 08:34.550 Knowledge can be added or removed instantly without retraining the model. 08:34.990 --> 08:41.350 Fine tuning may still be useful for style adaptation or task specialization, but it should not be the 08:41.350 --> 08:44.270 primary method for injecting factual knowledge. 08:44.750 --> 08:48.110 In most cases, Rag should be implemented first. 08:48.630 --> 08:53.870 The rule of thumb for engineers is simple use Rag before fine tuning. 08:54.310 --> 09:00.990 Rag provides flexibility, traceability, and control, all of which are essential for production AI 09:01.030 --> 09:01.750 systems. 09:02.430 --> 09:10.110 This final slide reinforces a critical message Rag is not optional for production or enterprise AI systems. 09:10.550 --> 09:12.350 As emphasized throughout the deck. 09:12.590 --> 09:17.910 Accuracy, trust and compliance are non-negotiable in real world deployments. 09:18.470 --> 09:25.550 Rag dramatically reduces hallucinations by grounding responses in retrieved verifiable sources. 09:25.990 --> 09:31.470 It enables auditability and traceability, which are required in regulated industries. 09:31.990 --> 09:38.550 Most importantly, it gives organizations control over what knowledge AI systems access and how that 09:38.550 --> 09:39.630 knowledge is used. 09:40.230 --> 09:47.390 At its core, Rag combines three elements llms for language, external knowledge for facts, and system 09:47.390 --> 09:49.950 level control for safety and governance. 09:50.590 --> 09:56.670 This combination transforms Llms from impressive demos into reliable business systems. 09:57.150 --> 09:59.870 The final insight is simple but powerful. 10:00.230 --> 10:04.750 Rag is the bridge between AI potential and enterprise reality. 10:05.190 --> 10:08.870 Without it, llms remain risky and unreliable. 10:09.190 --> 10:14.270 With it, they become trustworthy, scalable, and ready for production use.