<?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 Tanusha Kona on Medium]]></title>
        <description><![CDATA[Stories by Tanusha Kona on Medium]]></description>
        <link>https://medium.com/@tanushakona?source=rss-94686ea815d8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*fAoIy7iLf0BUfECrENEtOw.jpeg</url>
            <title>Stories by Tanusha Kona on Medium</title>
            <link>https://medium.com/@tanushakona?source=rss-94686ea815d8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 01:12:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tanushakona/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[All about the Jargons — Embeddings, Vector DB, Semantic Search — part 2]]></title>
            <link>https://medium.com/@tanushakona/all-about-the-jargons-embeddings-vector-db-semantic-search-part-2-41460c1f64e3?source=rss-94686ea815d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/41460c1f64e3</guid>
            <dc:creator><![CDATA[Tanusha Kona]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 17:13:42 GMT</pubDate>
            <atom:updated>2026-06-10T17:26:50.289Z</atom:updated>
            <content:encoded><![CDATA[<p>I hope you had a chance to read <a href="https://medium.com/@tanushakona/all-about-the-jargons-rag-llm-part-1-8df7c0c5a626">Part1 </a>of this series, where we explored some of the foundational GenAI concepts, including Large Language Models (LLMs), Retrieval-Augmented Generation (RAG). In this article, we’ll take the next step and dive deeper into the building blocks that make RAG systems work effectively. We’ll cover: Embeddings, Vector DB and Semantic Search.</p><h3>What Are Embeddings?</h3><p>Computers do not understand human language directly. So text is converted into numbers. These numerical representations are called as <em>Embeddings.</em></p><p>Example:</p><pre>&quot;Data Engineer&quot;</pre><p>becomes something like:</p><pre>[0.21, -0.55, 0.88, ...]</pre><p>These numbers represent the meaning of the sentence.</p><p>Why Are Embeddings Important?</p><p>Because similar meanings end up close together mathematically.</p><p>For example: “Machine Learning” and “AI Developer” will have similar embeddings. But :</p><pre>&quot;Pizza Recipe&quot;</pre><p>will be very far away. This enables semantic understanding.</p><h3>What is a Vector Database?</h3><p>A Vector Database stores embeddings.</p><p>Examples:</p><ul><li>ChromaDB</li><li>Pinecone</li><li>FAISS</li><li>Weaviate</li></ul><p>Instead of storing normal rows and columns like SQL databases, vector DBs store numerical meaning representations.</p><h3>Traditional Database vs Vector Database</h3><p>Traditional databases work like this:</p><pre>SELECT * <br>FROM jobs<br>WHERE skill = &#39;Python&#39;</pre><p>This is exact keyword matching. But Vector Databases work differently. They search based on meaning.</p><h3>Example</h3><p>Suppose the job description contains:</p><pre>&quot;Experience building AI pipelines&quot;</pre><p>But your resume says:</p><pre>&quot;Developed machine learning workflows&quot;</pre><p>A traditional database may miss this match. A vector database understands they are semantically similar.</p><h3>What is Semantic Search?</h3><p>Semantic Search means: <em>Searching by meaning instead of exact keywords. </em>This is one of the most powerful concepts in modern AI systems.</p><p>Traditional Search : If we search “ Data Engineer” ,we only get exact matches.</p><p><strong>Semantic Search</strong></p><p>If we search:</p><pre>&quot;AI workflow developer&quot;</pre><p>the system may retrieve:</p><ul><li>machine learning engineer</li><li>AI engineer</li><li>LLM developer</li></ul><p>because it understands context and meaning.</p><h3>How Semantic Search Works</h3><pre>Text<br>   ↓<br>Embeddings<br>   ↓<br>Vector Similarity<br>   ↓<br>Closest Meaningful Matches</pre><p>Usually this uses:</p><ul><li>cosine similarity</li><li>nearest neighbour search</li></ul><h3>Bringing Everything Together</h3><p>Now let’s see how all these concepts combine inside our AI Resume Tailor project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TAJlwXcezYjijpYwmPt-MQ.png" /></figure><h3>Final Thoughts</h3><p>When I first heard terms like:</p><ul><li>RAG</li><li>Vector DB</li><li>Semantic Search</li></ul><p>I assumed they were highly advanced AI research concepts.</p><p>But in reality:</p><ul><li>LLM = the brain</li><li>RAG = giving context</li><li>Embeddings = converting meaning into numbers</li><li>Vector DB = storing meanings</li><li>Semantic Search = finding similar meanings</li></ul><p>Once you understand this flow, modern AI architectures become much easier to understand. You can build these systems yourself using:</p><ul><li>LangChain</li><li>Ollama</li><li>Hugging Face</li><li>Streamlit</li><li>Open-source models</li></ul><p>In the next article, we’ll build a complete GenAI project together from scratch using LangChain and AI Agents. See you in the next part — where we move from concepts to code and start building!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=41460c1f64e3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[All about the Jargons ! — RAG, LLM — part 1]]></title>
            <link>https://medium.com/@tanushakona/all-about-the-jargons-rag-llm-part-1-8df7c0c5a626?source=rss-94686ea815d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/8df7c0c5a626</guid>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[rags]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Tanusha Kona]]></dc:creator>
            <pubDate>Thu, 28 May 2026 11:45:13 GMT</pubDate>
            <atom:updated>2026-05-29T15:28:06.507Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lU5wzoxa5ko5hG7ZI54z9g.png" /></figure><p>Where do I even start? What should I learn first? I kept breaking my head trying to understand the GenAI ecosystem. How do people catch up so quickly? Which YouTuber should I follow? Which Medium articles are actually useful? Which documentation should I read?</p><p>Honestly, I was completely confused and overwhelmed.</p><p>Relax. I was in the exact same place not too long ago. Think of it like cooking. You decide to try a new recipe even though you know you’re not a great cook. The first attempt may not look perfect, and half the time you’re just experimenting without knowing whether it’ll even taste good.</p><p>But eventually, you still eat it — because that’s how you learn.<br> (Unless laziness wins and you order food online instead. That’s an exception.)</p><p>Learning GenAI feels very similar.<br> At first, all the terms — LangChain, RAG, Vector Databases, Embeddings, LLMs — sound intimidating. But once you start building small things and experimenting consistently, the pieces slowly begin to connect. Let’s start.</p><h3>Understanding LLMs, RAG, Vector Databases, and Semantic Search — With a Simple AI Resume Builder Example</h3><p>When I first started learning Generative AI, I constantly heard terms like:</p><ul><li>LLMs</li><li>RAG</li><li>Vector Databases</li><li>Embeddings</li><li>Semantic Search</li></ul><p>Honestly, it all sounded intimidating.</p><p>But once I understood how these concepts work together in a real-world project, everything started making sense. So in this article, let’s understand these concepts using a simple project:</p><blockquote><em>Building an AI Resume Tailor that customises resumes according to job descriptions.</em></blockquote><h3>1. What is an LLM?</h3><p>LLM stands for <strong>Large Language Model</strong>.</p><p>Examples include: GPT , Llama, Gemini , Claude ,Mistral</p><p>An LLM is basically the “brain” that can: understand text, generate responses , summarise content, answer questions .Think of it like an extremely intelligent text prediction engine.</p><p>Example : suppose we ask:</p><pre>&quot;Rewrite my resume for a AI Engineer role&quot;</pre><p>The LLM analyses: your resume , the job description, important keywords and generates an improved version.</p><h4>But Here’s the Problem…</h4><p>The LLM does NOT automatically know: your resume, your company data or latest job postings. It only knows what was provided during training or in the prompt.</p><p>That’s where RAG comes in.</p><h3>2. What is RAG?</h3><p>RAG stands for: <em>Retrieval-Augmented Generation</em></p><p>It sounds complex, but the idea is simple: Instead of asking the LLM to answer blindly, we first retrieve relevant information and give it to the model. RAG is like giving an intelligent person access to the right notes and documents before asking them to answer a question.</p><h4>Real-Life Analogy</h4><p>Suppose you need to know the leave policy for the contractors of your company . The LLM itself does NOT know the company’s internal HR policies.</p><p>So the system:</p><ol><li>Searches company documents</li><li>Retrieves relevant HR policy pages</li><li>Sends those documents to the LLM</li><li>The LLM generates the final answer</li></ol><p>That is RAG. RAG Workflow in Our Resume Project</p><p>Suppose the user asks:</p><blockquote><em>“Am I a good fit for this AI Engineer role?”</em></blockquote><p>Without RAG:</p><ul><li>the AI may give generic advice</li></ul><p>With RAG:</p><ul><li>system first retrieves relevant sections from your resume</li><li>extracts matching skills from job description</li><li>sends them to the LLM</li></ul><p>Then the AI gives:</p><ul><li>personalized suggestions</li><li>tailored resume updates</li><li>skill gap analysis</li></ul><p>Now that we understand what RAG is and why it matters, the next step is understanding how AI systems actually “search by meaning.” In the next article, I’ll continue this journey using simple real-life analogies to explain:</p><ul><li>Semantic Search</li><li>Embeddings</li><li>Vector Databases</li><li>LangChain</li><li>AI Agents</li></ul><p>and how all of them work together inside modern Generative AI applications. Because once these concepts click together, the entire Gen AI ecosystem starts making much more sense.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8df7c0c5a626" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Partitioning vs Clustering — interview focused answer]]></title>
            <link>https://medium.com/@tanushakona/partitioning-vs-clustering-interview-focused-answer-bfea70741288?source=rss-94686ea815d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/bfea70741288</guid>
            <category><![CDATA[clustering]]></category>
            <category><![CDATA[spark]]></category>
            <category><![CDATA[query-optimization]]></category>
            <category><![CDATA[bigquery]]></category>
            <category><![CDATA[data-engineering]]></category>
            <dc:creator><![CDATA[Tanusha Kona]]></dc:creator>
            <pubDate>Thu, 14 May 2026 14:10:42 GMT</pubDate>
            <atom:updated>2026-05-14T14:10:42.980Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*SUcBWUHNrD3_5u_N.jpg" /></figure><p>The key idea is:</p><ul><li>Partitioning narrows data first at a coarse level (for example, by date).</li><li>Clustering narrows data again inside that partition at a fine level (for example, by country, event_type).</li></ul><p>A simple mental model imagine your data is a massive pile of invoices. If you just throw them all in one drawer. When you want to find a specific invoice, you have to check <strong>every single page </strong>:</p><ol><li>Partition = choose the right drawer.</li><li>Cluster = find the right folder inside that drawer.</li></ol><p>Why clustering helps after partitioning:</p><ul><li>Even one partition (like one day or month) can still be huge.</li><li>Clustering stores similar values together, so fewer data blocks are read.</li><li>This improves filters, and often helps JOIN and GROUP BY performance because related values are colocated.</li></ul><p>When to use what:</p><ol><li>Partitioning only<br>Use when almost every query filters by date/time and there is no stable second filter.</li><li>Clustering only<br>Use when there is no strong time filter, but queries repeatedly filter/group by columns like region, country, product, customer segment.</li><li>Both partition + cluster<br>Best for very large time-series tables where users filter by date plus one or more dimensions (for example date + country).</li><li>BigQuery example pattern:</li></ol><ul><li>Partition by event date</li><li>Cluster by frequently filtered dimensions like country and event_type</li></ul><p>This gives two levels of pruning:</p><ol><li>Partition pruning (skip irrelevant dates)</li><li>Block pruning via clustering (skip irrelevant blocks within selected partitions)</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bfea70741288" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Partition Pruning vs Predicate Pushdown — Interview focused answer]]></title>
            <link>https://medium.com/@tanushakona/partition-pruning-vs-predicate-pushdown-interview-focused-answer-754833a04f51?source=rss-94686ea815d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/754833a04f51</guid>
            <category><![CDATA[query-optimization]]></category>
            <category><![CDATA[spark]]></category>
            <category><![CDATA[sql]]></category>
            <category><![CDATA[partitioning]]></category>
            <category><![CDATA[optimization]]></category>
            <dc:creator><![CDATA[Tanusha Kona]]></dc:creator>
            <pubDate>Thu, 14 May 2026 12:51:36 GMT</pubDate>
            <atom:updated>2026-05-14T12:51:36.512Z</atom:updated>
            <content:encoded><![CDATA[<h3>Partition Pruning:</h3><ul><li>Skips entire partitions based on partition boundaries/metadata.</li><li>Works at the partition level (coarse-grained).</li><li>Example: table partitioned by month; filter for June → skip Jan, Feb, Mar, Apr, May, Jul, etc. partitions entirely.</li><li>Decides which partitions to read before opening them.</li></ul><h3>Predicate Pushdown:</h3><ul><li>Pushes filter conditions (WHERE clause) down to the data source layer.</li><li>Works at the row/column level within partitions (fine-grained).</li><li>Applies filtering logic at the source instead of reading all data into Spark first.</li><li>Uses column metadata (min/max values) to skip data chunks that don’t match the filter.</li><li>Reduces data transfer from source to Spark.</li></ul><p>Simple analogy:</p><ol><li>Partition pruning: “Skip entire filing cabinets that don’t contain June files.”</li><li>Predicate pushdown: “Within the June cabinet, only pull out invoices &gt; $1000, not all invoices.”</li></ol><p><strong>How they work together:</strong></p><pre>Query: SELECT * FROM sales WHERE month = &#39;June&#39; AND amount &gt; 1000<br><br>1. Partition pruning (coarse): <br>   Skip Jan, Feb, Mar, Apr, May, Jul... cabinets entirely.<br><br>2. Predicate pushdown (fine):<br>   In the June cabinet, apply amount &gt; 1000 at the file level <br>   (using min/max metadata) before reading rows into Spark.<br><br>3. Result:<br>   Only matching June rows with amount &gt; 1000 are transferred to Spark.</pre><p><strong>Key difference:</strong></p><ul><li>Pruning: partition-level elimination (before opening partitions).</li><li>Pushdown: row-level filtering at source (within partitions).</li></ul><p>Both reduce data scanned, but at different granularities.</p><p><strong>Why helpful:</strong></p><ol><li>Reduces data scanned (less I/O).</li><li>Faster query execution (fewer partitions/rows to process).</li><li>Lower network costs (less data transfer).</li><li>Cheaper cloud bill (pay for data scanned, not stored).</li></ol><p><strong>Best practice:</strong><br>Use both together:</p><ol><li>Prune partitions first.</li><li>Then pushdown filters on remaining data.</li></ol><p>This maximises performance and cost savings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T0B7gGlHOmV0QCUvG_2WjQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=754833a04f51" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hive Interview Based Key Points]]></title>
            <link>https://medium.com/@tanushakona/hive-interview-based-key-points-d2a13f78db4e?source=rss-94686ea815d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2a13f78db4e</guid>
            <category><![CDATA[hive]]></category>
            <category><![CDATA[hadoop]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[interview-preparation]]></category>
            <dc:creator><![CDATA[Tanusha Kona]]></dc:creator>
            <pubDate>Fri, 17 May 2024 19:20:44 GMT</pubDate>
            <atom:updated>2024-05-17T19:26:10.429Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><strong>Hive Interview Based Key Points</strong></blockquote><blockquote><strong>Key features of Hive</strong>:</blockquote><ul><li>Hive works on the HDFS, and you can query into the HDFS with an SQL-like querying language using Hive.</li><li>It supports multiple read operations but a single write function.</li><li>It supports the Hive Query Language, which is similar to the SQL, in order to query into the HDFS. However, all the queries are converted into MapReduce jobs only.</li><li>It stores all the information of the tables, including data types, relational structure and schema, into its metastore.</li><li>It has a variety of aggregate functions and a special user-defined function (UDF) feature.</li><li>It supports batch processing to process the data after each time period.</li><li>Along with the feature of querying into the HDFS using the HQL, Hive supports several applications that are written in Java or Python and can also interact with the HDFS using Hive drivers. Therefore, Hive is highly extensible.</li></ul><blockquote><strong>The architecture of Hive</strong>: <br>Hive consists of the following three components:</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*wvlA7jQmlmnXVDk_.png" /></figure><p><strong>Hive client</strong>: It contains CLI and web interface to interact with Hive using the Hive Querying language.</p><p><strong>Hive core</strong>: This is an integral part of the Hive architecture. It consists of the following three aspects:</p><p><strong>1.Metastore</strong>: It stores the schema information of the data that has been read using the Hive queries from the HDFS. It uses Derby RDBMS to store the schema of the tables.</p><p><strong>2.Hive QL process engine</strong>: It checks the syntax of the HQL queries and sends it to the metastore in order to derive the schema information for the query.</p><p><strong>3.Execution engine</strong>: It converts the HQL queries into MapReduce jobs in order to read the data from the HDFS.</p><p><strong>Hadoop cluster</strong>: It consists of HDFS and MapReduce. HDFS is a file storage system that stores the files in a distributed manner, whereas MapReduce is used to query into the HDFS.</p><blockquote>Difference Between RDBMS &amp; Hive:</blockquote><p><strong>RDBMS vs Hive</strong>: One of the main differences between RDBMS and Hive is that Hive works on ‘schema on read’ while RDBMS works on ‘schema on write’. Apart from this, Hive takes longer to compute, whereas RDBMS requires less time to query. The data size is larger in Hive than in RDBMS.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d2a13f78db4e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>