Fine Tuning RAG Vs Agentic AI

Key Takeaways

rack fine-tuning prompt engineering different ways. So important some professionals competing but these three are different solutions for three different problems. So engineering stepbyep. This is your AA coach John [music] compensation up to 1.2 2 cr CTC for exceptional candidates. The number one highly demandable skill set that is a engineer back in the 2026 retali agents [music] the AWS GCPs or Azour Five plus years of software engineering experience up to 1.2 in the particular session 1.2 two scorecated but in the particular session less than 20 minutes concept scriptural diagram. So in the session starts or GCP Professionals platforms high level. Okay. Okay. So now the very first point Azour a architecture overview. So high level overview. So first different for different for example Google drive object storage services for example information multiple services. So the process multiple services large language model most of the companies Microsoft rightoft modal So first multiple multiple sources datyp different different file for Enterprise data is stored, process and enriched. Then process then GP generates intelligence response response large language model. So GP does not know your data because policies out of the box GP has no knowledge of your internal policies or documents or business processes. And the question how do we bridge this particular gap language modeling each solving a different problemat fine tuning rack. So the very first services. Five servicesion 24 hours. Next steps a run two a companies we have another company as wellies Okay. What if you may ask? What is your refund policy? What is the delivery policy? Can I return the products after 15 days? services 4.1 Google learning platform Google what is provided policy. As of my latest update, there is no widely recognized publicly available information about the company or service. language model but the model is doesn't does not know your business. Uh the first problem solution technical dictionar language complete Retund available within seven days. So response multiple storage units in the particular existing for best for stable knowledge. Okay. Rarely changing data. So what is this particular now? Customize the model. Okay. Train with domain specific examples. Knowledge becomes part of the model. Okay. Specialized use cases industry terminology consistent tone and domain specific reasoning built into the model itself. So example policy the model learns your HR policies policy rules internal terminology during the training already existing is nothing but what existing modeling teaches the model by updating its weights during the training knowledge is baked in permanently Okay. Expensive retraining every time the data changes. So from month by month or five to seven days expensive retraining every time hours of hours or days of training knowledge becomes outstanded immediately after the training model cannot handle dynamic or real data real example. Policy changes 7 to 30 days model 7 days until fully retained. Fully training cycle required just to finish one fact problem. What if data changes every day? Bottom line cannot keep up. We need a solution that uses the latest information without retraining. bottom. Fine tuning works only for stable rarely changing knowledge problem. not for dynamic data which comes for the rescue. Retrieval augmented generation comes for the rescue. So basic policies 30 days placement support available multiple documents. Okay. Multiple documents. Multiple documents documents for documents. So first a search multiple services model context protocol vision One of the search questions relatable or close in the container because context because always up to date. Okay. How it work? First search. Okay. So, rag gives knowledge. Sorry. Rack gives GPT access to knowledge. Ra gives GPE instead of forcing it to memorize everything during the training. Right. Forement different. Okay. types of different rack and retrie information from the documents. Right? Find refund policy instantly. Okay. Answers answer knowledge based questions. Super. But what it cannot do? Perform actions or execute workflows. Okay. Send the email to the customer. For example, find a policy in the particular customer. Rag finds the policy but rag cannot send the email. It only answers not act problems. So the gap is businesses need a that can act not just answers. rag stand. We are entering into this agent system and the agent system that a that reasons plans retrieves and executes information a system. So first send the refund policy to customers or all customers multiple functions or okay functionol for concept. Okay. policy policy in 30 days. workflow. Email successfully. Last stepcess. Okay. Act as a strategist and take this particular policies and send it to this customer. Okay. agent multiple search dat Microsoft Orchestration. Hey, I want to prepare one document. Task 25 years experienced document expert or documentation expert respation Agent understands, decides and acts full business automation. So agent combines reasoning agent system. Okay. Stable knowledge rarely changing information. Okay. high level understand 1.2 2.2 to see. Do not directly aim for that. But the conception language model description next If you have any kind of questions we can clarify if you wanted to make a study. Okay. Right. So first point if you're just dropping your comments I will get to Because preparation at least we are spending some time. So just take two minutes time and let us know in the comment section what value you have learned from this. First drop your takeway points learning points and then ask your questions so we will definitely plan here. Okay, this is your a John signing off. Bye. [music] [music] >> [music] >> Heat.