<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Abhinaya Jeyakumar  on Medium]]></title>
        <description><![CDATA[Stories by Abhinaya Jeyakumar  on Medium]]></description>
        <link>https://medium.com/@abhisakthika?source=rss-28a258c8c8da------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*OTOedV0ZKHKnNuor</url>
            <title>Stories by Abhinaya Jeyakumar  on Medium</title>
            <link>https://medium.com/@abhisakthika?source=rss-28a258c8c8da------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:13:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@abhisakthika/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[RAG From Zero: The Mental Model I Wish I Had as a Beginner]]></title>
            <link>https://medium.com/linkit-intecs/rag-from-zero-the-mental-model-i-wish-i-had-as-a-beginner-3a54faaac820?source=rss-28a258c8c8da------2</link>
            <guid isPermaLink="false">https://medium.com/p/3a54faaac820</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[rags]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Abhinaya Jeyakumar ]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 08:51:26 GMT</pubDate>
            <atom:updated>2026-09-25T08:51:26.973Z</atom:updated>
            <content:encoded><![CDATA[<h4>How I learned why RAG exists, what chunks and embeddings actually do, and how an LLM finds the right information before answering.</h4><h4>RAG Without the Buzzwords - Part 1 of 2</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*broppYEvF9dWWa9Op-CrNQ.png" /><figcaption>RAG overview. AI-generated by the author using ChatGPT.</figcaption></figure><p>For the past few months, I kept hearing one term again and again whenever people discussed Large Language Models:</p><p><strong>RAG.</strong></p><p>At first, I thought learning RAG would simply mean learning one new AI concept.</p><p>Then I started searching.</p><p>Suddenly, I was seeing terms such as <strong>embeddings, chunks, vector databases, semantic search, retrievers, reranking, Agentic RAG, GraphRAG, Self-RAG, and CRAG</strong>.</p><p>Instead of becoming clearer, the topic became more confusing.</p><p>The problem was not that each concept was extremely difficult. I was simply trying to learn them in the wrong order.</p><p>I was learning the names before understanding the problem they were created to solve.</p><p>So I went back to the beginning and asked one question:</p><p><strong>Why does an LLM need RAG in the first place?</strong></p><p>Once I understood that, the other pieces started connecting.</p><p>This article is the beginner-friendly explanation I wish I had when I first started learning RAG.</p><h3>Why I Wanted to Understand RAG</h3><p>As an Information Technology undergraduate, I had already worked with normal software systems where data usually comes from APIs, databases, or backend services.</p><p>LLM-based applications introduced a different question.</p><p>If I ask an AI system something about information that it has never seen before, <strong>where does the answer come from?</strong></p><p>For example, imagine building an AI assistant for a university.</p><p>A student asks:</p><p><strong>“Can I submit my assignment after the deadline?”</strong></p><p>Suppose the latest student handbook says:</p><p><strong>“Assignments may be submitted up to three days late. A 10% penalty applies for each late day.”</strong></p><p>The answer exists.</p><p>But a general-purpose LLM does not automatically have access to that private university handbook.</p><p>This was the first idea I needed to understand before learning RAG.</p><h3>What an LLM Knows and What It Can See Are Different</h3><p>Large Language Models learn patterns and information during training.</p><p>But that does not mean they automatically know:</p><ul><li>your company’s latest policy,</li><li>a private university document,</li><li>today’s internal announcement,</li><li>information inside your personal PDFs,</li><li>the latest database records,</li><li>or every document created after their training.</li></ul><p>There is another important concept here: <strong>context</strong>.</p><p>Context is the information available to the model while it is producing its current response.</p><p>So there is an important difference between:</p><p><strong>what the model learned during training</strong></p><p>and</p><p><strong>what information we give the model right now.</strong></p><p>If our LLM has never seen the latest student handbook and we simply ask:</p><p><strong>“Can I submit my assignment late?”</strong></p><p>it may not have enough reliable information to answer.</p><p>It might say it does not know.</p><p>It might give a general answer based on similar information.</p><p>Or it may generate something that sounds reasonable but is not actually supported by our university’s policy.</p><p>That is where the real problem starts.</p><h3>Why Not Give the LLM Every Document?</h3><p>My next thought was simple.</p><p>If the LLM does not know the handbook, why not just give it the entire handbook?</p><p>For one small document, that might actually work.</p><p>But imagine the university assistant has access to:</p><ul><li>hundreds of announcements,</li><li>academic regulations,</li><li>module guides,</li><li>examination policies,</li><li>timetables,</li><li>student handbooks,</li><li>and years of archived documents.</li></ul><p>A student asks one question about a late assignment.</p><p>Most of that information is completely irrelevant.</p><p>The student does not need the whole university archive.</p><p>They need the <strong>right piece of information</strong>.</p><p>This became the first mental shortcut that made RAG easier for me:</p><p><strong>Don’t give the LLM everything. Find what matters first.</strong></p><p>And that is basically where RAG begins.</p><h3>What Is RAG?</h3><p><strong>RAG stands for Retrieval-Augmented Generation.</strong></p><p>The name sounds complicated, but the basic idea is surprisingly simple.</p><p>We can break it into three parts.</p><p><strong>Retrieval</strong></p><p>Find information that is relevant to the user’s question.</p><p><strong>Augmented</strong></p><p>Add that retrieved information to the context given to the LLM.</p><p><strong>Generation</strong></p><p>Allow the LLM to generate its response using that information.</p><p>Without RAG, we can think of the process as:</p><p><strong>Question → LLM → Answer</strong></p><p>With RAG:</p><p><strong>Question → Find relevant information → Give it to the LLM → Generate the answer</strong></p><p>The original RAG research by Patrick Lewis and colleagues combined generation with information retrieved from an external knowledge source rather than relying only on knowledge stored inside the model.</p><p>For me, the easiest definition is:</p><p><strong>RAG finds useful external information first and gives it to an LLM before asking the model to answer.</strong></p><h3>The Open-Book Exam Analogy</h3><p>The easiest way I found to remember RAG is to think about an exam.</p><p>Imagine two students.</p><p>The first student has to answer every question using only what they remember.</p><p>The second student is allowed to use their notes.</p><p>However, the second student does not read every textbook from beginning to end for every question.</p><p>They first find the page related to the question.</p><p>Then they read it.</p><p>Then they answer.</p><p>That is very similar to what RAG does.</p><p><strong>Without RAG</strong></p><p>Question → Existing knowledge → Answer</p><p><strong>With RAG</strong></p><p>Question → Find relevant information → Read the evidence → Answer</p><p>So my simple mental model is:</p><p><strong>RAG turns a closed-book-style LLM workflow into something closer to an open-book workflow.</strong></p><p>However, there is one important catch.</p><p>Having access to the textbook is not enough.</p><p>The student still needs to find the <strong>correct page</strong>.</p><p>If the wrong page is selected, the final answer can still be wrong.</p><p>That becomes very important later when we look at retrieval quality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*esDHU0yprOZo0W4H9KbeKQ.png" /><figcaption>Closed-book vs open-book LLM workflow. AI-generated by the author using ChatGPT.</figcaption></figure><h3>Step 1: Breaking Documents into Chunks</h3><p>Let’s return to our university handbook.</p><p>Imagine it contains 200 pages.</p><p>Searching the entire document as one huge block would not be very useful.</p><p>Instead, the document can be divided into smaller meaningful pieces.</p><p>These are called <strong>chunks</strong>.</p><p>For example:</p><p><strong>Chunk 1 — Attendance Policy</strong></p><p>Students must maintain at least 80% attendance.</p><p><strong>Chunk 2 — Assignment Policy</strong></p><p>Assignments may be submitted up to three days late. A 10% penalty applies for each late day.</p><p><strong>Chunk 3 — Medical Extension Policy</strong></p><p>Students experiencing illness may request an extension with an approved medical certificate.</p><p>A chunk is simply a smaller piece of a larger document that can be searched and retrieved more easily.</p><p>This sounds simple, but chunking can strongly affect RAG quality.</p><h3>Why Chunk Size Matters</h3><p>Consider this sentence:</p><p><strong>“Students experiencing illness may request an assignment extension by submitting an approved medical certificate.”</strong></p><p>Imagine splitting it into:</p><p><strong>Chunk A:</strong> Students experiencing illness may…</p><p><strong>Chunk B:</strong> …request an assignment extension…</p><p><strong>Chunk C:</strong> …by submitting an approved medical certificate.</p><p>If our retrieval system returns only Chunk B, we know that an extension can be requested.</p><p>But we have lost two important details:</p><ul><li>who qualifies for the extension,</li><li>and what evidence is required.</li></ul><p>At the same time, making chunks extremely large is not automatically better.</p><p>One chunk might then contain information about assignments, examinations, attendance, library rules, and medical leave together.</p><p>The meaning becomes less focused.</p><p>So chunking requires a balance:</p><p><strong>Small enough to retrieve accurately, but large enough to preserve meaning.</strong></p><h3>Beginner Shortcut: There Is No Perfect Chunk Size</h3><p>You may see tutorials recommending values such as:</p><ul><li>300 tokens,</li><li>500 tokens,</li><li>1,000 tokens.</li></ul><p>These can be useful starting points, but they are not universal rules.</p><p>Instead of asking:</p><p><strong>“What is the best chunk size?”</strong></p><p>a better question is:</p><p><strong>“What kind of document am I splitting, and what kinds of questions will users ask?”</strong></p><p>A source-code repository, legal agreement, FAQ collection, and student handbook may need very different chunking strategies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ek3gklTMqFugtYDF6CnkGg.png" /><figcaption>Breaking a student handbook into meaningful chunks. AI-generated by the author using ChatGPT.</figcaption></figure><h3>Step 2: Understanding Embeddings</h3><p>Once I understood chunks, I reached the next confusing sentence:</p><p><strong>“Generate embeddings and store them in a vector database.”</strong></p><p>This is usually where beginner explanations become much more technical.</p><p>But embeddings make sense if we first understand the problem they solve.</p><p>Imagine a document contains this sentence:</p><p><strong>“Employees may work remotely.”</strong></p><p>A user asks:</p><p><strong>“Can staff work from home?”</strong></p><p>The exact words are different.</p><p><strong>Employees</strong> is not the same word as <strong>staff</strong>.</p><p><strong>Work remotely</strong> is not the same phrase as <strong>work from home</strong>.</p><p>But their meanings are very similar.</p><p>If our retrieval system only searched for identical words, it might miss useful information.</p><p>We therefore need a way for computers to compare <strong>meaning</strong>.</p><p>That is where embeddings become useful.</p><h3>What Is an Embedding?</h3><p>An embedding model converts text into a numerical representation.</p><p>For example:</p><p><strong>“work from home”</strong></p><p>might become something conceptually similar to:</p><p><strong>[0.21, -0.47, 0.81, …]</strong></p><p>That list of numbers is called a <strong>vector</strong>.</p><p>The individual numbers are not normally something we manually interpret.</p><p>What matters is the relationship between different vectors.</p><p>Texts with similar meanings can receive representations that are relatively close to each other in what is called <strong>embedding space.</strong></p><p>So:</p><p><strong>“work from home”</strong></p><p>and</p><p><strong>“remote work”</strong></p><p>should be much more semantically related than:</p><p><strong>“database transaction isolation”</strong></p><p>Current vector-search systems work with this idea by converting documents and queries into numeric vectors and comparing them using similarity measures.</p><p>The definition that helped me most is:</p><p><strong>Embedding = representing meaning with numbers so that computers can compare information.</strong></p><h3>Beginner Shortcut: Translate the Jargon</h3><p>When you hear:</p><p><strong>“We embed the documents.”</strong></p><p>mentally translate it into:</p><p><strong>“We convert the text into numerical representations that allow us to compare meaning.”</strong></p><p>The terminology suddenly becomes much less intimidating.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C_-QEvPfTeR1GCrPa76PnQ.png" /><figcaption>Conceptual semantic similarity visual. AI-generated by the author using ChatGPT.</figcaption></figure><p><em>Note: The positions are illustrative and do not represent real embedding coordinates.</em></p><h3>Step 3: Searching by Meaning</h3><p>Now suppose our handbook contains three chunks:</p><ul><li>Attendance Policy</li><li>Assignment Policy</li><li>Medical Extension Policy</li></ul><p>The student asks:</p><p><strong>“Can I submit my assignment late?”</strong></p><p>The question can also be converted into an embedding.</p><p>The system can then compare the question representation with the stored document representations.</p><p>Conceptually, the results may look like:</p><p><strong>Assignment Policy — highly relevant</strong></p><p><strong>Medical Extension Policy — somewhat relevant</strong></p><p><strong>Attendance Policy — less relevant</strong></p><p>The system can then retrieve the strongest matches.</p><p>This is the basic idea behind <strong>semantic search</strong>.</p><p>Semantic search tries to find information based on meaning, rather than depending only on identical words.</p><h3>Does This Mean Keyword Search Is Bad?</h3><p>No.</p><p>This is another thing I initially misunderstood.</p><p>Imagine searching for:</p><p><strong>ICT3012</strong></p><p>or</p><p><strong>INC-2489</strong></p><p>or a specific product code.</p><p>In these cases, the exact characters may matter more than semantic similarity.</p><p>Keyword search can be extremely useful.</p><p>That is why more advanced retrieval systems often combine different search methods instead of treating semantic search as the answer to every problem.</p><p>The important lesson is:</p><p><strong>Search by meaning when meaning matters. Search exactly when exact terms matter.</strong></p><p>We will return to this idea in Part 2.</p><h3>Step 4: Where the Vector Database Fits</h3><p>At this point, each chunk may have:</p><ul><li>its original text,</li><li>an embedding,</li><li>and additional information about where it came from.</li></ul><p>That additional information is called <strong>metadata</strong>.</p><p>For example:</p><p><strong>Text:</strong> Assignments may be submitted up to three days late.</p><p><strong>Source:</strong> Student Handbook 2026</p><p><strong>Section:</strong> Assignment Policy</p><p><strong>Year:</strong> 2026</p><p><strong>Embedding:</strong> numerical vector representing the text</p><p>We then need some way to store or index these vectors and search them efficiently.</p><p>This is where a <strong>vector store</strong> or <strong>vector database</strong> can be useful.</p><p>A vector store stores embedded data and supports similarity search over those vectors.</p><h3>Beginner Shortcut: RAG Is Not a Vector Database</h3><p>This is an important distinction.</p><p>A vector database can be <strong>one component</strong> inside a RAG system.</p><p>It is not RAG itself.</p><p>Retrieval could also use:</p><ul><li>keyword search,</li><li>a normal database,</li><li>an external API,</li><li>a website,</li><li>a search engine,</li><li>or a knowledge graph.</li></ul><p>A <strong>retriever</strong> is the broader component responsible for returning useful documents for a query. It does not have to be a vector store.</p><p>So when someone says:</p><p><strong>“We use a vector database.”</strong></p><p>you know something about how their system searches information.</p><p>You still do not know their complete RAG architecture.</p><h3>Step 5: Following One Question Through RAG</h3><p>Now all the individual terms start connecting.</p><p>Before users ask questions, we prepare the documents.</p><p><strong>Student Handbook</strong></p><p>↓</p><p><strong>Extract the text</strong></p><p>↓</p><p><strong>Split it into chunks</strong></p><p>↓</p><p><strong>Create embeddings</strong></p><p>↓</p><p><strong>Store or index those chunks</strong></p><p>This preparation stage is commonly called <strong>indexing</strong>.</p><p>Then the student asks:</p><p><strong>“Can I submit my assignment late?”</strong></p><p>The system:</p><ol><li>converts the question into a searchable representation,</li><li>searches for relevant information,</li><li>retrieves the Assignment Policy,</li><li>adds the retrieved policy to the model’s context,</li><li>and asks the LLM to answer.</li></ol><p>The LLM may effectively receive:</p><p><strong>Context:</strong> Assignments may be submitted up to three days late. A 10% penalty applies for each late day.</p><p><strong>Question:</strong> Can I submit my assignment late?</p><p>The final response could be:</p><p><strong>“Yes. According to the supplied policy, assignments may be submitted up to three days late, with a 10% penalty for each late day.”</strong></p><p>The most important point here is:</p><p><strong>The LLM did not need to permanently learn our university policy during training.</strong></p><p>The system retrieved the information when it was needed and supplied it as context.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Uj8ye0BA5bR9U8kG5RTAUg.png" /><figcaption>Simplified end-to-end RAG pipeline. AI-generated by the author using ChatGPT.</figcaption></figure><h3>The Complete RAG Pipeline</h3><p>At this point, I started seeing RAG as two simple stages.</p><h3>1. Prepare the Knowledge</h3><p>Documents → Chunks → Embeddings → Searchable Index</p><h3>2. Answer a Question</h3><p>Question → Retrieve Relevant Information → Add Context → LLM → Answer</p><p>That one mental model made the rest of the topic much easier.</p><p>Instead of seeing chunks, embeddings, vector stores, retrievers, and LLMs as unrelated technologies, I could finally see what role each one played.</p><h3>What Does Top-K Mean?</h3><p>Once you start reading more about RAG, you will quickly meet another technical-looking term:</p><p><strong>Top-K retrieval.</strong></p><p>Fortunately, this one is simple.</p><p>If:</p><p><strong>K = 3</strong></p><p>the system retrieves the top three results.</p><p>If:</p><p><strong>K = 5</strong></p><p>it retrieves five.</p><p>The difficult part is not understanding the term.</p><p>The difficult part is choosing a useful value.</p><p>If we retrieve 50 chunks when only three are relevant, we may fill the LLM’s context with unnecessary information.</p><p>If we retrieve only one chunk, we might miss another important piece of evidence.</p><p>So another RAG trade-off appears:</p><p><strong>Give the model enough evidence without giving it unnecessary noise.</strong></p><h3>What Is Reranking?</h3><p>Suppose our first search retrieves 20 potentially useful chunks.</p><p>Instead of sending all 20 directly to the LLM, the system can examine those candidates more carefully and reorder them according to their relevance.</p><p>That is called <strong>reranking</strong>.</p><p>The workflow might look like:</p><p><strong>Thousands of chunks → Fast retrieval → 20 candidates → Reranker → Best 4 → LLM</strong></p><p>I found this easier to remember using two questions.</p><p><strong>Retriever:</strong> Which documents look promising?</p><p><strong>Reranker:</strong> Which of these promising documents are actually the best?</p><p>And this reveals something important.</p><p>We did not replace RAG with a completely different technology.</p><p>We simply improved one stage of the original pipeline.</p><p>Remember that idea.</p><p>It becomes extremely useful in Part 2.</p><h3>The Retriever Is the Researcher. The LLM Is the Writer.</h3><p>This is another mental model I found useful.</p><p>The <strong>retriever</strong> searches for useful evidence.</p><p>The <strong>LLM</strong> reads that evidence and turns it into a clear answer.</p><p>So I think of them as:</p><p><strong>Retriever = Researcher</strong></p><p><strong>LLM = Writer</strong></p><p>This distinction is also extremely useful when something goes wrong.</p><h3>Why a Powerful LLM Can Still Give a Bad RAG Answer</h3><p>Imagine the student asks:</p><p><strong>“Can I submit my assignment late?”</strong></p><p>But the retriever gives the model:</p><p><strong>“Students must maintain at least 80% attendance.”</strong></p><p>The LLM may be extremely capable.</p><p>But it has been given the wrong evidence.</p><p>This led me to one of the most useful RAG debugging habits:</p><p><strong>When a RAG answer is bad, inspect what was retrieved before blaming the LLM.</strong></p><h3>Beginner Debugging Hack: Find Where the Answer Disappeared</h3><p>Instead of randomly changing models, ask these questions in order:</p><ol><li>Does the source contain the answer?</li><li>Was the document read or parsed correctly?</li><li>Is the answer inside a meaningful chunk?</li><li>Was that chunk retrieved?</li><li>Did it rank high enough?</li><li>Was it actually sent to the LLM?</li><li>Did the LLM use the evidence correctly?</li></ol><p>The first place where the answer becomes <strong>No</strong> tells you much more about the real problem.</p><p>For example:</p><p><strong>Source contains answer — Yes</strong></p><p><strong>Chunk contains answer — Yes</strong></p><p><strong>Retriever returned chunk — No</strong></p><p>Now changing the LLM probably will not solve the issue.</p><p>The retrieval stage needs attention.</p><p>This way of debugging RAG is much more useful than treating the entire system as one mysterious AI box.</p><h3>Three Things I Wish I Had Known Earlier</h3><h3>1. RAG Does Not Mean Training an LLM on Your PDF</h3><p>In a typical RAG workflow, information is retrieved when needed and placed into the model’s context.</p><p>That is different from <strong>fine-tuning</strong>, where the model itself is further trained.</p><p>My shortcut is:</p><p><strong>RAG mainly changes the context the model receives at query time.</strong></p><p><strong>Fine-tuning changes the model’s parameters and behaviour.</strong></p><p>We will look at that distinction properly in Part 2.</p><h3>2. RAG Does Not Mean Vector Database</h3><p>Vector search is common, but it is only one retrieval method.</p><p>The core idea is still:</p><p><strong>retrieve useful information → give it to the model → generate an answer.</strong></p><h3>3. RAG Does Not Guarantee Zero Hallucinations</h3><p>The wrong document may be retrieved.</p><p>An outdated document may appear.</p><p>Important evidence may be missing.</p><p>The model may misunderstand the retrieved text.</p><p>Or it may generate something that is not properly supported by the evidence.</p><p>RAG can improve grounding.</p><p>It does not make correctness automatic.</p><h3>You Do Not Need a Huge Project to Learn RAG</h3><p>One mistake I nearly made was thinking I needed a complicated framework, cloud infrastructure, and a large knowledge base before I could properly understand RAG.</p><p>You do not.</p><p>Start with only three pieces of information.</p><p><strong>Attendance Policy</strong></p><p>Students must maintain at least 80% attendance.</p><p><strong>Assignment Policy</strong></p><p>Assignments may be submitted up to three days late. A 10% penalty applies for each late day.</p><p><strong>Medical Extension Policy</strong></p><p>Students experiencing illness may request an extension with an approved medical certificate.</p><p>Then ask:</p><p><strong>“Can I submit my assignment late?”</strong></p><p>If your system finds the Assignment Policy and gives it to the LLM, you have already implemented the core idea.</p><p>Everything else can be added when there is an actual reason for it.</p><h3>What I Learned</h3><p>Learning RAG became much easier once I stopped trying to memorise every term separately.</p><p>I learned that:</p><ul><li><strong>chunks</strong> make large documents easier to retrieve from,</li><li><strong>embeddings</strong> represent text numerically so meaning can be compared,</li><li><strong>semantic search</strong> helps find information with similar meaning,</li><li><strong>vector stores</strong> are one way of supporting similarity search,</li><li><strong>retrievers</strong> are responsible for finding useful information,</li><li><strong>rerankers</strong> help improve the order of retrieved results,</li><li>and the <strong>LLM</strong> uses the retrieved evidence to generate the final response.</li></ul><p>But the biggest lesson was simpler than any of those definitions.</p><p><strong>The pipeline matters more than the buzzwords.</strong></p><p>Once I understood the basic pipeline, many advanced RAG terms stopped looking like completely different technologies.</p><h3>The Trick That Changed How I Learn New RAG Terms</h3><p>I kept hearing:</p><ul><li>Naive RAG</li><li>Advanced RAG</li><li>Modular RAG</li><li>Agentic RAG</li><li>GraphRAG</li><li>Self-RAG</li><li>Corrective RAG</li></ul><p>At first, I thought I needed to memorise seven separate architectures.</p><p>Then I realised that many RAG techniques are trying to improve, replace, or add something around the same basic pipeline.</p><p>Maybe the user’s query needs to be rewritten.</p><p>Maybe retrieval needs to use more than semantic similarity.</p><p>Maybe the retrieved evidence needs to be checked.</p><p>Maybe relationships between information matter.</p><p>Maybe the system should decide whether it needs another search.</p><p>That gave me one rule that I now use whenever I encounter another RAG term:</p><p><strong>Don’t memorise the name first. Find the box that changed, then ask why it changed.</strong></p><p>That is exactly where Part 2 begins.</p><h3>Conclusion</h3><p>RAG sounded complicated to me because I first encountered it through its tools and terminology.</p><p>Once I started from the actual problem, the idea became much simpler.</p><p>An LLM may not have access to the information needed for a particular question.</p><p>Instead of trying to put everything into the model, RAG retrieves the information that matters and gives that evidence to the LLM at the right time.</p><p>From there, the main workflow becomes easier to understand:</p><p><strong>Documents → Chunks → Embeddings → Retrieval → Context → LLM → Answer</strong></p><p>That foundation matters more than memorising every framework or database.</p><p>For me, understanding this basic pipeline was the point where RAG stopped looking like a collection of confusing AI terms and started looking like an understandable engineering system.</p><h3>Coming Next — Part 2</h3><h3>Why Are There So Many RAG Architectures? Find the Box That Changed</h3><p>In Part 2, I will use the same basic pipeline from this article to understand:</p><ul><li>Naive RAG</li><li>Advanced RAG</li><li>Modular RAG</li><li>Agentic RAG</li><li>GraphRAG</li><li>Self-RAG</li><li>Corrective RAG</li></ul><p>Instead of memorising each definition, we will ask two questions:</p><p><strong>Which part of basic RAG changed?</strong></p><p><strong>What problem was that change trying to solve?</strong></p><p>I will also compare <strong>RAG, fine-tuning, and long-context models</strong>, and look at why a more complicated RAG architecture is not automatically a better one.</p><h3>References and Further Reading</h3><p><strong>Lewis, P. et al. (2020).</strong> <a href="https://arxiv.org/abs/2005.11401"><em>Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.</em></a> Advances in Neural Information Processing Systems 33.</p><p><strong>Gao, Y. et al. (2023).</strong> <em>R</em><a href="https://arxiv.org/abs/2312.10997"><em>etrieval-Augmented Generation for Large Language Models: A Survey.</em></a> This survey discusses the evolution from Naive RAG to Advanced RAG and Modular RAG.</p><p><strong>LangChain Documentation.</strong> <a href="https://docs.langchain.com/oss/python/langchain/knowledge-base"><em>Build a Semantic Search Engine with LangChain.</em></a> Useful further reading on embeddings, vector stores, similarity search, and retrievers.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a54faaac820" width="1" height="1" alt=""><hr><p><a href="https://medium.com/linkit-intecs/rag-from-zero-the-mental-model-i-wish-i-had-as-a-beginner-3a54faaac820">RAG From Zero: The Mental Model I Wish I Had as a Beginner</a> was originally published in <a href="https://medium.com/linkit-intecs">LinkIT</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Turning Customer Data into Retention Decisions: My Journey into Churn Prediction]]></title>
            <link>https://medium.com/linkit-intecs/turning-customer-data-into-retention-decisions-my-journey-into-churn-prediction-4e17bfabd16e?source=rss-28a258c8c8da------2</link>
            <guid isPermaLink="false">https://medium.com/p/4e17bfabd16e</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[power-bi]]></category>
            <category><![CDATA[student-project]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Abhinaya Jeyakumar ]]></dc:creator>
            <pubDate>Fri, 28 Aug 2026 06:18:09 GMT</pubDate>
            <atom:updated>2026-08-28T06:18:09.541Z</atom:updated>
            <content:encoded><![CDATA[<h4>How I used Python, SQL, machine learning, and Power BI to analyse customer churn and identify high-risk customer groups.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EkB2ngeXARUKdCHFosdnQA.png" /><figcaption>Customer Retention Intelligence Dashboard created by the author using Power BI.</figcaption></figure><p>As an Information Technology undergraduate, most of my earlier projects were closer to software development. In those projects, the main focus was usually on building features, connecting the front end with the back end, handling database operations, and making sure the system worked correctly.</p><p>When I started working on this data science project, I realised that the thinking process was different.</p><p>In software development, we usually ask questions like:</p><p>Can the user log in?<br>Can the user add, update, or delete records?<br>Is the API working correctly?<br>Is the user interface responsive?</p><p>In a data science project, the questions are different:</p><p>What does the data tell us?<br>Is the data clean enough to use?<br>What patterns can we find?<br>Can we predict something useful?<br>How can the result support a business decision?</p><p>To understand this workflow practically, I built an end-to-end data science and business intelligence project called <strong>Customer Retention Intelligence Dashboard</strong>.</p><p>The project focuses on analysing customer churn, predicting churn using machine learning, and presenting the final findings through a Power BI dashboard.</p><p>Customer churn means customers leaving a company or stopping the use of a service. For a telecommunication company, churn is an important business problem because losing customers directly affects revenue. If a company can identify customers who are more likely to leave, it can take action earlier and improve customer retention.</p><h3>Why I Chose This Project</h3><p>I wanted to build a project that was not only technical but also connected to a real business problem.</p><p>Customer churn prediction was a good choice because it helped me learn the complete data science workflow, including data cleaning, analysis, visualisation, machine learning, SQL, dashboard creation, and business recommendations.</p><p>This project also helped me understand how data science is different from normal software development. Instead of only building features, I had to understand the data, find patterns, train models, and explain the results in a meaningful way.</p><h3>Project Overview</h3><p>The main goal of this project was to analyse telecom customer data and identify which customers are more likely to churn.</p><p>The project answers questions such as:</p><ul><li>What percentage of customers left the company?</li><li>Which contract types have the highest churn?</li><li>Which payment methods are linked with higher churn?</li><li>Do monthly charges affect churn?</li><li>Does customer tenure affect churn?</li><li>Which customer group has the highest churn risk?</li><li>Can machine learning help predict customer churn?</li></ul><p>By answering these questions, the project gives both technical results and business insights.</p><h3>Tools and Technologies Used</h3><p>For this project, I used:</p><ul><li>Python</li><li>Pandas</li><li>NumPy</li><li>Matplotlib</li><li>Seaborn</li><li>Scikit-learn</li><li>Joblib</li><li>MySQL</li><li>Power BI</li><li>Git and GitHub</li></ul><p>Python was used for data cleaning, exploratory data analysis, visualisation, machine learning preparation, model training, and feature importance analysis.</p><p>MySQL was used to perform SQL-based analysis.</p><p>Power BI was used to build the final dashboard and present the findings in a clear visual format.</p><p>GitHub was used to maintain the project repository and documentation.</p><h3>Dataset Used</h3><p>I used the <strong>Telco Customer Churn dataset</strong>, which is publicly available on Kaggle and is based on IBM Sample Data Sets.</p><p>Dataset link: <a href="https://www.kaggle.com/datasets/blastchar/telco-customer-churn">Telco Customer Churn Dataset on Kaggle</a></p><p>The dataset contains customer details such as gender, senior citizen status, partner and dependent details, tenure, phone service, internet service, contract type, payment method, monthly charges, total charges, and churn status.</p><p>The original dataset contained <strong>7043 rows and 21 columns</strong>.</p><p>After cleaning, the dataset contained <strong>7032 rows and 21 columns</strong>.</p><p>I removed 11 rows because the total charges field had missing values after converting it into numeric format.</p><h3>Step 1: Understanding the Data</h3><p>The first step was to load the dataset using Python and Pandas.</p><p>I checked the basic structure of the dataset by looking at the first few rows, the number of rows and columns, column data types, basic statistics, and churn value counts.</p><p>These checks helped me understand what information was available in the dataset, how many customers were included, and how many customers had churned or stayed.</p><p>At this stage, I noticed that the total charges field was stored as text instead of a numeric value. Since this field represents a payment amount, it had to be converted before analysis and machine learning.</p><p>This step taught me that before building any machine learning model, understanding the dataset is very important.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/727/1*gAaM9NUnV8P46_BPE_-4gg.png" /><figcaption>Initial data inspection showing the dataset preview, shape, important data types, and churn count.</figcaption></figure><h3>Step 2: Data Cleaning</h3><p>Data cleaning was one of the most important parts of this project.</p><p>The main cleaning task was converting the total charges field from text to numeric format. During this conversion, 11 values became missing because they were not valid numeric values.</p><p>Since 11 rows were very small compared to the full dataset, I removed those rows.</p><p>After cleaning, I saved the cleaned dataset as a new CSV file and used it for further analysis.</p><p>This step helped me understand that data science is not only about training models. If the data is not cleaned properly, the final results can become inaccurate or misleading.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/405/1*gU1_0t2ZMwh4PFMF5K9g3g.png" /><figcaption>Cleaning summary after handling missing total charges.</figcaption></figure><h3>Step 3: Exploratory Data Analysis</h3><p>After cleaning the data, I performed exploratory data analysis, also known as EDA.</p><p>EDA helped me understand the main patterns inside the data before applying machine learning.</p><p>I analysed overall churn count, overall churn percentage, churn by gender, churn by contract type, churn by internet service, churn by payment method, average monthly charges by churn, and average tenure by churn.</p><p>The overall churn rate was <strong>26.58%</strong>.</p><p>This means around one-fourth of the customers in the dataset had left the company.</p><p>One of the clearest findings was that month-to-month contract customers had a much higher churn rate than customers with one-year or two-year contracts.</p><p>The churn rate by contract type was:</p><ul><li>Month-to-month contract: <strong>42.71%</strong></li><li>One-year contract: <strong>11.28%</strong></li><li>Two-year contract: <strong>2.85%</strong></li></ul><p>This showed that customers with long-term contracts were more stable, while month-to-month customers were more likely to leave.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/392/1*JdyjdH1RGGAtWyAc_5_K4A.png" /><figcaption>Cleaning summary after handling missing total charges.</figcaption></figure><h3>Step 4: Data Visualisation</h3><p>After the basic analysis, I created charts using Matplotlib and Seaborn.</p><p>I created visualisations for churn count, churn by contract type, churn by internet service, churn by payment method, monthly charges by churn, and tenure by churn.</p><p>These charts helped me understand the patterns more clearly.</p><p>For example, it was easier to visually see that churn was higher among month-to-month contract customers, fibre optic internet customers, electronic check payment users, customers with higher monthly charges, and customers with lower tenure.</p><p>This step helped me understand why visualisation is important in data science. Tables give exact numbers, but charts make patterns easier to understand and explain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/707/1*uszMFh0Qe2Ppn7BkLQTKSA.png" /><figcaption>Contract type visualisation showing higher churn among month-to-month customers.</figcaption></figure><h3>Step 5: Preparing the Data for Machine Learning</h3><p>Before training machine learning models, the dataset had to be prepared.</p><p>The customer ID field was removed because it is only an identifier. It does not help the model understand churn behaviour.</p><p>The target field was churn. I converted it into numeric values, where “Yes” became 1 and “No” became 0. This was needed because machine learning models work with numbers.</p><p>I also converted categorical information such as contract type, internet service, and payment method into numeric columns using one-hot encoding.</p><p>In the dataset, these values were stored as text, but machine learning models need numbers to make predictions.</p><p>After this step, the machine-learning-ready dataset had <strong>7032 rows and 31 columns</strong>.</p><p>This step helped me understand that preparing data for machine learning is a separate and important part of the workflow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/487/1*vKnY2iY8nwJ0B0l1DBxDfA.png" /><figcaption>Machine learning data preparation summary.</figcaption></figure><h3>Step 6: Training Machine Learning Models</h3><p>For the prediction part, I trained and compared three machine learning models:</p><ul><li>Logistic Regression</li><li>Decision Tree</li><li>Random Forest</li></ul><p>After comparing the models, the final selected model was <strong>Logistic Regression with Scaling</strong>.</p><p>The final model performance was:</p><ul><li>Accuracy: <strong>80.38%</strong></li><li>Precision: <strong>64.76%</strong></li><li>Recall: <strong>57.49%</strong></li><li>F1 Score: <strong>60.91%</strong></li></ul><p>I selected Logistic Regression with Scaling because it gave the best F1 score among the tested models.</p><p>In this project, F1 score was important because churn prediction needs a balance between precision and recall. A company wants to identify customers who may leave, but the predictions should also be reasonably reliable.</p><p>This step helped me understand that model selection is not only about choosing the highest accuracy. The evaluation metric should match the problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/687/1*Vwfw0Y9ocqaLM5NS4bs-Bg.png" /><figcaption>Model comparison summary.</figcaption></figure><h3>Step 7: Understanding Feature Importance</h3><p>After training the final model, I analysed which features were important for churn prediction.</p><p>Some important churn-related features were tenure, monthly charges, fibre optic internet service, total charges, contract type, electronic check payment, online security, and tech support.</p><p>This helped me move beyond just prediction. Instead of only asking, “Will this customer churn?”, I could also ask, “What factors are connected with churn?”</p><p>For example, customers with lower tenure were more likely to churn. Customers with higher monthly charges also showed higher churn risk. Month-to-month contracts and electronic check payment methods were also strongly connected with churn.</p><p>This part was important because business users need reasons behind the prediction, not just the final output.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/1*ROEjtkbWjUQwIn-kaiJmYQ.png" /><figcaption>Feature importance analysis showing key churn-related factors.</figcaption></figure><h3>Step 8: SQL Analysis Using MySQL</h3><p>I also used MySQL to analyse the cleaned dataset.</p><p>The cleaned CSV file was imported into MySQL, and I wrote SQL queries to confirm the patterns found in Python.</p><p>Using SQL, I analysed overall churn rate, churn by contract type, churn by payment method, churn by internet service, average monthly charges by churn, average tenure by churn, and the highest-risk customer group.</p><p>This step was useful because SQL is widely used in real-world data analysis and business intelligence work.</p><p>It also helped me confirm that the insights found using Python were consistent with the SQL results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/913/1*ZXZlg9mJH9YXZvqh6Al-nw.png" /><figcaption>SQL analysis summary by contract type.</figcaption></figure><h3>Step 9: Building the Power BI Dashboard</h3><p>The final stage of the project was creating a Power BI dashboard.</p><p>The dashboard includes total customers, churned customers, retained customers, overall churn rate, high-risk churn rate, churn by contract type, churn by payment method, churn by internet service, average monthly charges by churn, average tenure by churn, and the highest-risk customer group recommendation.</p><p>The dashboard made the project easier to understand from a business point of view. Instead of only showing code, CSV files, and model results, the dashboard presents the key insights visually.</p><p>For me, this was one of the most useful parts of the project because it connected technical analysis with business communication.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EkB2ngeXARUKdCHFosdnQA.png" /><figcaption>Final Power BI dashboard created to present churn insights and high-risk customer groups.</figcaption></figure><h3>Key Findings</h3><p>The main findings from this project were:</p><h3>1. Overall churn rate was 26.58%</h3><p>Around one-fourth of customers in the dataset had churned.</p><h3>2. Month-to-month customers had the highest churn rate</h3><p>Month-to-month contract customers had a churn rate of <strong>42.71%</strong>, while two-year contract customers had only <strong>2.85%</strong> churn.</p><p>This suggests that long-term contracts are linked with better customer retention.</p><h3>3. Electronic check users had higher churn</h3><p>Customers using electronic check payment had a churn rate of <strong>45.29%</strong>.</p><p>This may indicate that payment method behaviour is connected with churn risk.</p><h3>4. Fibre optic customers had higher churn</h3><p>Fibre optic internet customers had a churn rate of <strong>41.89%</strong>.</p><p>This could point to possible issues related to pricing, service expectations, or customer satisfaction.</p><h3>5. Churned customers had higher monthly charges</h3><p>The average monthly charge for churned customers was <strong>74.44</strong>, while the average monthly charge for retained customers was <strong>61.31</strong>.</p><p>This suggests that customers with higher monthly charges may need more attention.</p><h3>6. Churned customers had lower tenure</h3><p>The average tenure of churned customers was <strong>17.98 months</strong>, while retained customers had an average tenure of <strong>37.65 months</strong>.</p><p>This shows that newer customers were more likely to leave.</p><h3>7. The highest-risk group had 60.37% churn</h3><p>The highest-risk customer group was made up of customers with a month-to-month contract, fibre optic internet service, and an electronic check payment method.</p><p>This group had a churn rate of <strong>60.37%</strong>.</p><p>This is much higher than the overall churn rate, so this group should be prioritised for retention campaigns.</p><h3>Business Recommendations</h3><p>Based on the analysis, a telecom company can take the following actions:</p><ul><li>Focus retention campaigns on month-to-month customers.</li><li>Encourage customers to move to one-year or two-year contracts.</li><li>Investigate fibre optic service quality or pricing concerns.</li><li>Provide personalised offers to customers with high monthly charges.</li><li>Encourage automatic payment methods such as credit card or bank transfer.</li><li>Pay more attention to newer customers during their early months.</li><li>Prioritise the highest-risk customer group for customer retention campaigns.</li></ul><p>This part helped me understand the practical value of data science. The final goal is not only to build a model but also to support decision-making.</p><h3>What I Learned</h3><p>This project helped me understand the full data science workflow.</p><p>I learned how to load and inspect a dataset, clean incorrect and missing values, perform exploratory data analysis, create visualisations, prepare data for machine learning, train and compare models, evaluate model performance, understand feature importance, use SQL for analysis, create a Power BI dashboard, and convert technical results into business recommendations.</p><p>One of the biggest lessons I learned is that machine learning is only one part of a data science project. Data cleaning, analysis, visualisation, and business understanding are equally important.</p><h3>How This Project Was Different from Software Development</h3><p>As an IT undergraduate, I found this project different from a typical software development project.</p><p>In software development, the output is usually a working feature or system. In this project, the output was insight.</p><p>Instead of building screens and APIs, I had to understand the data, find patterns, train a model, and explain the results in a way that could support business decisions.</p><p>This project helped me understand how data science connects technical work with real-world decision-making.</p><h3>Conclusion</h3><p>The Customer Retention Intelligence Dashboard project gave me practical experience in data science and business intelligence.</p><p>By working on this project, I learned how raw customer data can be cleaned, analysed, prepared for machine learning, and finally presented through a dashboard to support business decision-making.</p><p>The project helped identify important churn patterns and high-risk customer groups. It also showed how data can be used to make practical business recommendations.</p><p>For me, this project was not just about completing a portfolio project. It was a learning experience that helped me understand how data science works in a real-world business problem.</p><h3>Project Repository</h3><p>GitHub Repository:<br><a href="https://github.com/abhi190503/customer-retention-intelligence-dashboard">Customer Retention Intelligence Dashboard on GitHub</a></p><h3>Image and Dataset Credits</h3><p>Image credit: Dashboard screenshot created by the author using Power BI.</p><p>Dataset credit: Telco Customer Churn Dataset by BlastChar on Kaggle, based on IBM Sample Data Sets.<br>Dataset source: <a href="https://www.kaggle.com/datasets/blastchar/telco-customer-churn">Telco Customer Churn Dataset on Kaggle</a><br>IBM sample data reference: <a href="https://community.ibm.com/community/user/businessanalytics/blogs/steven-macko/2019/07/11/telco-customer-churn-1113">IBM Telco Customer Churn Sample Data</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4e17bfabd16e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/linkit-intecs/turning-customer-data-into-retention-decisions-my-journey-into-churn-prediction-4e17bfabd16e">Turning Customer Data into Retention Decisions: My Journey into Churn Prediction</a> was originally published in <a href="https://medium.com/linkit-intecs">LinkIT</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Passed Our Code Review Using AI — But I Still Questioned Myself]]></title>
            <link>https://medium.com/@abhisakthika/we-passed-our-code-review-using-ai-but-i-still-questioned-myself-6af73690f043?source=rss-28a258c8c8da------2</link>
            <guid isPermaLink="false">https://medium.com/p/6af73690f043</guid>
            <category><![CDATA[students]]></category>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Abhinaya Jeyakumar ]]></dc:creator>
            <pubDate>Fri, 22 May 2026 18:45:55 GMT</pubDate>
            <atom:updated>2026-05-22T18:45:55.789Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mk7mtMB509igOmaF3W5YNw.png" /></figure><p><em>Thoughts of a second-year IT undergraduate navigating AI dependency, academic pressure, and self-doubt.</em></p><h3>The Code Review That Stayed in My Mind</h3><p>A few weeks ago, our software project went through a code review.</p><p>As second-year IT undergraduates, we were dealing with many things at once — unfamiliar technologies, tight deadlines, bugs we barely understood, and the pressure of making everything work under time constraints.</p><p>We did manage to complete it successfully.</p><p>And honestly, AI tools played a major role in helping us get through it.</p><p>At that moment, it felt like we had survived something difficult.</p><p>But after everything settled, I didn’t feel the relief I expected.</p><p>Instead, I felt a quiet question forming in my mind.</p><h3>“Am I Actually Learning, or Just Using AI to Survive?”</h3><p>That thought stayed longer than I expected.</p><p>AI tools are extremely helpful. They speed up learning, help debug errors, and guide us when we’re stuck.</p><p>But slowly, a concern started building underneath the convenience.</p><p>Am I really improving my skills?</p><p>Or am I becoming dependent on tools without realising it?</p><p>I think many IT students today silently face this question, especially in the AI era where everything feels one prompt away.</p><h3>The Pressure That Doesn’t Always Look Visible</h3><p>Not long after that, we finished our mid exams.</p><p>The days leading up to them were intense.</p><p>Assignments, projects, and exams all came together in a way that left very little space to breathe.</p><p>There were days that felt mentally overwhelming from morning until night.</p><p>And over time, even physical stress started showing signs — fatigue, overthinking, and constant pressure.</p><p>It made me wonder if this level of stress is something we are expected to just “normalise” in university life.</p><h3>Some Students Seem Calm — And That Makes It More Confusing</h3><p>One thing I keep noticing is how differently students handle pressure.</p><p>Some people seem completely calm — they laugh during busy weeks, submit work on time, and still perform well without visible stress.</p><p>Meanwhile, I tend to overthink everything — deadlines, performance, future internships, and whether I’m doing enough.</p><p>It makes me ask myself:</p><p>Do they really feel less pressure, or are they just better at handling it?</p><p>I don’t think there’s a clear answer.</p><h3>Loving IT, But Struggling With the Process</h3><p>What makes this more confusing is that I genuinely enjoy my degree.</p><p>I chose IT because I liked technology — building systems, solving problems, and understanding how things work.</p><p>And academically, things are going well.</p><p>But I slowly realised something important:</p><p>Enjoying a subject is very different from going through the process of studying it.</p><p>And sometimes, that process feels heavier than expected.</p><h3>University Life Didn’t Feel Like This in My Mind</h3><p>Before joining university, I imagined a balanced life.</p><p>Studying, learning new things, enjoying time with friends, exploring interests, and growing at a steady pace.</p><p>But reality often feels different.</p><p>Most days feel like a cycle of:<br> assignments → deadlines → exams → pressure → repeat.</p><p>And somewhere in between, personal time and enjoyment start fading into the background.</p><h3>The Fear That Comes With Internships</h3><p>As internships get closer, another layer of pressure starts forming.</p><p>There’s excitement, of course — stepping into real industry work is something I look forward to.</p><p>But there is also fear.</p><p>If university already feels overwhelming at times, what will real workplace pressure feel like?</p><p>Am I prepared enough?</p><p>Am I skilled enough?</p><p>These are questions that don’t always have clear answers, but they stay in the background quietly.</p><h3>Maybe We Are All Learning How to Carry Pressure</h3><p>One thing I’m slowly realising is this:</p><p>Maybe being “fine” doesn’t mean being stress-free.</p><p>Maybe it means learning how to carry pressure while still moving forward.</p><p>Maybe the students who look calm aren’t unaffected — they’ve just learned different ways to manage what we all feel.</p><h3>Final Thoughts</h3><p>Right now, I don’t have perfect answers.</p><p>I’m still learning how to balance:</p><ul><li>ambition and rest,</li><li>learning and pressure,</li><li>AI tools and real understanding,</li><li>success and mental peace.</li></ul><p>But I do know one thing.</p><p>I am not the only student thinking this way.</p><p>And maybe that itself is something worth talking about more openly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6af73690f043" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>