<?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 Raunak Regmi on Medium]]></title>
        <description><![CDATA[Stories by Raunak Regmi on Medium]]></description>
        <link>https://medium.com/@raunakregmi?source=rss-4f7637b426d3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*hhg14bdk4RJQlRQBHEXATw.jpeg</url>
            <title>Stories by Raunak Regmi on Medium</title>
            <link>https://medium.com/@raunakregmi?source=rss-4f7637b426d3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 05:46:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@raunakregmi/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[The Fable 5 Shutdown: What Happens When an AI Refuses to Help?]]></title>
            <link>https://ai.plainenglish.io/when-an-ai-refuses-to-help-what-the-fable-5-shutdown-says-about-the-future-of-ai-safety-b82f035bf094?source=rss-4f7637b426d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/b82f035bf094</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[claude-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[Raunak Regmi]]></dc:creator>
            <pubDate>Sat, 13 Jun 2026 05:31:49 GMT</pubDate>
            <atom:updated>2026-06-13T19:30:19.305Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>A few days ago, I came across a statement from Anthropic that immediately caught my attention. What happened?</blockquote><p>The company stated that under instruction from the government of the United States, the ability of foreign citizens to access the most advanced version, Fable 5 and Mythos 5, would temporarily halt. Anthropic stated that there was a jailbreak process involved which may be a clever bypass to certain safeguards of the model.</p><p>This decision came rather abruptly ‘just today’. The team at Anthropic did not agree with this reasoning, citing the fact that these claimed vulnerabilities existed only within certain parameters and were not different from what other advanced AI systems could do to a larger extent.</p><p>While opening the page and reading through the declaration, I found myself thinking differently with different experience.</p><p>Not with Fable.</p><p>With Claude.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OxH0FPd3wj1dr9PNy0JuBg.png" /><figcaption>Claude’s Statement on Banning Fable 5 and Mythos 5 taken from <a href="https://www.anthropic.com/news/fable-mythos-access">Anthropic</a>.</figcaption></figure><p>Claude’s official news: <a href="https://www.anthropic.com/news/fable-mythos-access">https://www.anthropic.com/news/fable-mythos-access</a></p><p>The Frustration of Being Told “No”</p><p>Being part of the developer and student community, I use and experiment with AI models on a regular basis understanding their capabilities and limitations.</p><p>There have been times when I have tried using Claude Opus for academic purposes, only to find out that it refuses some demands. Regardless of how I framed my queries or changed the circumstances around my questions, it just led to arguments still there were demands it wouldn’t fulfill.</p><p>At the time, surely, I was frustrated.</p><p>As a user, one would obviously seek to use the most advanced model out there for their benefit. With the need for information and help should come easily. It is frustrating when a model rejects a command from the user.</p><p>Frustrated, I switched back to Sonnet and continued what I had to fulfill (my work).</p><p>At the time, the experience didn’t seem like much.</p><p>Today, it feels far more relevant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DtouRvL6V7rJ7DD2XhMbSg.jpeg" /><figcaption>When Claude Refused To Do My Work. Taken from my Conversation with <a href="https://claude.ai/login?returnTo=/new?">Claude</a>.</figcaption></figure><h3>The Hidden Challenge of Frontier AI</h3><p>Most discussions around artificial intelligence focus on capabilities.</p><ul><li>Which model is smarter?</li><li>Which model writes better code?</li><li>Which model performs best on benchmarks?</li><li>Which model achieves the highest scores?</li></ul><p>Obviously, these are important, but they ignore challenges even harder to solve.</p><p>Models are this powerful, but how make it safer?</p><p>True that, building intelligence is difficult. But,</p><p>Building such intelligence consistently remaining on its limit is hell of a work.</p><p>Stronger the model, developers, writers, or users like me assume they are capable of doing more and more complicated ask. In addition to us making sure that these ability of powerful models won’t be misused.</p><p>This created a tension which all the frontier AI companies are facing today.</p><p>Everywhere is a problem,</p><p>Users complain to safeguard being too strict.</p><p>If such safeguards are proved to be inadequate, regulators, scientists, and governments become worried.</p><p>Maintaining this balance is even more difficult.</p><h3>Why Jailbreaking Matters?</h3><p>“Jailbreak” doesn’t always mean completely taking the control of a certain model, which, many people surely assume.</p><p>This isn’t the reality, its a common subtle situation.</p><p>Jailbreak, generalized, simply is a way to convince a model to ignore few imposed rules to it. Jailbreaks tend to be often more specific and strongly requires certain conditions to succeed.</p><p>Most challenging part isn’t this, its AI safety isn’t judged by its design.</p><p>Rather, judged by its potential consequences.</p><p>Vulnerability might appear minor, but we can’t ignore tha fact that it could be adapted or expanded into far more serious and harmful extent, with it being a concern.</p><p>According to the message by Anthropic, the weaknesses revealed to them were not of value and comparable to other weaknesses in any open-source model.</p><p>Anyway, it does not really make much difference which of the sides one takes.</p><p>Concerns regarding safety can trigger government intervention just like this, when huge frontier AI systems are currently operating at a scale</p><p>That alone signals a major shift in how advanced AI is viewed.</p><h3>The Capability Race Is Becoming a Safety Race</h3><p>Coming up to this day, AI industry are categorized as a race for bringing out powerful models.</p><ul><li>Bigger models.</li><li>Larger context windows.</li><li>More reasoning ability.</li><li>Better performance.</li></ul><p>With these advancements happening regularly, another competition emerges along their side.</p><p>The main race, the race for bulding systems worthy of trust.</p><p>The next “crazy” breakthrough at this era will not come from a model scoring higher on benchmarks.</p><p>Rather, it may come from a model like “Opus 4.8” which can be relied on resisting manipulative decisions and that still is powerful and useful.</p><p>But, isn’t it much harder engineering problem? Many people don’t realize this.</p><p>Forget LLMs and models for now,</p><p>Us, humans, we ourselves struggle to maintain a consistent boundary under pressure and wouldn’t it be unrealistic to expect that from a machine to perfectly maintaining that under every possible scenario?</p><p>Yet that is increasingly becoming the expectation.</p><h3>The Irony of AI Safety</h3><p>The interesting part of this discussion is the irony remaining at the center.</p><p>The same safeguards that frustrated users complain about are often the safeguards that enable them to use the model at first place.</p><p>When Claude refused some of my requests, I viewed it as a limitation.</p><p>Looking back, that refusal was evidence that the system was actively enforcing rules designed by its creators.</p><p>The more capable a model becomes, the more important those refusals become.</p><p>A model that never says “no” might feel more useful in the short term.</p><p>But that could lead it to become significantly more dangerous.</p><h3>Looking Ahead</h3><p>My personal opinion, the future of AI will never be determined completely by someone building the most intelligent model.</p><p>Intelligence alone is never enough for long.</p><p>The companies that succeed long term will likely be those that can balance everything we talked about “capability, safety, usability, and trust”.</p><p>The recent suspension of Fable 5 and Mythos 5 is a reminder that frontier AI is entering a new phase.</p><p>The conversation is no longer just about what models can do.</p><p>What’s increasingly becoming more relevant is what they must do, what they must avoid doing, and whether that boundary can hold its ground against efforts to breach it, “jailbreak”.</p><p>In the expanding, ever evolving era of artificial intelligence, theres a question that remains within our minds:</p><p>Will the most important feature of future AI models be their intelligence, or their ability to know when not to use it?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b82f035bf094" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/when-an-ai-refuses-to-help-what-the-fable-5-shutdown-says-about-the-future-of-ai-safety-b82f035bf094">The Fable 5 Shutdown: What Happens When an AI Refuses to Help?</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Claims to Know Nepali Culture — Or Is It Just Guessing?]]></title>
            <link>https://ai.plainenglish.io/ai-claims-to-know-nepali-culture-or-is-it-just-guessing-755d0db4e428?source=rss-4f7637b426d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/755d0db4e428</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-evaluation]]></category>
            <category><![CDATA[benchmarking]]></category>
            <dc:creator><![CDATA[Raunak Regmi]]></dc:creator>
            <pubDate>Sun, 22 Feb 2026 17:05:42 GMT</pubDate>
            <atom:updated>2026-02-23T17:48:29.757Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI can speak your cultural language — But can it grasp the heart of Nepali culture and its traditions?</h4><p><strong>Small Mistake, a Big Gap</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ryzWc1_9UeIi52sihjAnsw.png" /><figcaption>AI Unable to Understand Local Nepali Language. Created using <a href="https://gemini.google/overview/image-generation/">Gemini 2.5 (w/ Nano Banana)</a> by Author.</figcaption></figure><p>Imagine this, you are lazy, asking AI to do everything. You ask AI to greet your elder uncle in Nepali culture who has just arrived.</p><p>It responds with ease, even confidently: “हे, के छ?” [“Hey, what’s up?”]</p><p>Wait — thats not right? Your uncle isn’t your casual friend like it responded and most definitely you address your elders with respect which is built in <strong>words, gestures and your tone</strong> in Nepali culture.</p><p>With greetings like “नमस्ते” [<em>“Namaste”</em>], hierarchical words like “आउनु होस्” [“Please Come”].</p><p>That answer it gave? Completely Wrong.</p><p>AI can speak Nepali no-doubt, but can it understand <strong>social hierarchy</strong>, understand core sarcasm, <strong>indigenous kinships</strong>? Let’s answer it.</p><h3>The Illusion of Multilingual AI</h3><p>LLMs undoubtedly are grammatically accurate, fluent and versatile with linguistics. AI easily and accurately can speak Nepali, translate it and talk in it. Modern AI is impressive, GPTs can write many languages.</p><p>On surface, AI can speak tons of languages — but the real catch comes here:</p><p><strong>cultural meaning</strong> <strong>≠ language fluency.</strong></p><p>Consider kinships like, “भान्जीको बुहारी” [<em>“Bhanji ko buhari”</em>], ‘Trans: Grand-niece-in-law’. These have their own call-outs like “नातिनी बुहारी<strong>” </strong>[<em>“Natini buhari”</em>].</p><p>Social hierarchies like, “दाजु” [<em>“Daju”</em>], “Trans: Elder Brother”, carry their own respect.</p><p>Festivals, <em>jatras </em>like Dashain, <em>Indrajatra </em>(इन्द्रजात्रा) aren’t just simple dates on calendar. They are rituals, family codes holding some tradition and customs followed from ages.</p><p>Idioms (<em>ukkhan/tukka</em>) like, “आँखा चिम्लनु” literally translates to “closing eyes”. Is it even right?</p><p>Ancestors brought up these <em>ukkhan/tukka </em>as a way of sarcasm, humor which needs an indigenous understanding and emotions — they aren’t just literal translations.</p><p>AI might get the grammar right, but these cultural linguistic layer of Nepali languages? Often wrong and inaccurate.</p><h3>Nepal Is Invisible in Training Data</h3><p>Why is this a major issue?</p><p>Why does this happen?</p><p>The simple answer is: <strong>Low-resource </strong>Nepali languages online</p><p>Mostly, AI is trained under the data that is available online. There certainly is a vast bias where English dominates it. Nepali content exists, but it’s a tiny fraction compared to English. Much of Nepali context is based on their <strong>daily-living, oral traditions, life-style, local festivals, local languages, daily life conversations</strong> and many more, and they rarely make it online.</p><p>Based on data online, AI learns the language — but not the culture behind it.</p><p>For our benchmarking, we collected data related with core Nepali culture which was rare online:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1D9o3p4QBA_4spRJDfHQpA.png" /><figcaption>Example Of The Dataset. By <a href="http://raunakregmi.medium.com/">Author</a>.</figcaption></figure><p>This is the part of solution to the problem Nepal faces.</p><h3><strong>Where AI Fails Locally</strong></h3><p>We tested this with concrete examples,</p><ul><li><strong>Cultural Etiquette:</strong> AI can describe the general etiquette to be followed, but it mixes up with what actually happens in a Nepali household.</li><li><strong>Agriculture:</strong> During crop harvesting, seeding farmers rely on the weather and Nepalese calendars where AI guesses the western dates.</li><li><strong>Kinship &amp; Social Hierarchy:</strong> Nepalese family have their unspoken rules for who bows, who eats first, what relationship is between the family members.</li><li><strong>Idioms/Poverbs (<em>ukkhan/tukka)</em>: </strong>Nepali often use them while speaking sarcastically and in a certain humorous way, which often doesn’t make a literal sense.</li><li><strong>Gastronomy: </strong>The local kitchen culture of Nepal is totally different for what spices on which food and the way of cooking, which is deeply anchored with Nepali culture only.</li></ul><p>These everyday situations where Nepali culture is attached to. AI fails here — it’s funny — quite misleading too.</p><h3>Why Benchmarks Matter</h3><p>A simple analogy,</p><p>Benchmarks are exams for AI. Without exams, a student could memorize multiple linguistic phrases, sound smart right?</p><p>But can they bring it to <strong>real-life interactions</strong>? That is what benchmarks tests. Even some family relationships are still confusing for the people around — when will AI learn it?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HI5MOGLR7rZNlkQxVO_k6Q.png" /><figcaption>Nepali family tree (Spouse). Image from <a href="https://nepalgo.de/vocabulary/vocabulary-family-and-kinship/">Nepalgo</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KWjs7Jg9wa-T78jkGUP21g.jpeg" /><figcaption>Nepali family tree (Immediate Family). Image from <a href="https://nepalgo.de/vocabulary/vocabulary-family-and-kinship/">Nepalgo</a></figcaption></figure><p>For AI, benchmarks measures wether AI can obtain true common-sense rather than just vocabulary. Without it, AI seems smart — but surfacely.</p><h3>What We Built</h3><p>Can AI really understand Nepali culture?</p><p>Arising from this question, we created a Nepali commonsense benchmark (NeCCo) where we tested among locally collected datas in sparse categories we talked about above — not to brag, but to show AI its blindspots that Nepali language isn’t just about translation or vocabulary — it’s more common-sense.</p><p>Well, we got some “must to check out” results:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3ERUvHGd9_cly9oCFdKmjA.png" /><figcaption>Comparative Accuracy across Categories. By <a href="http://raunakregmi.medium.com/">Author</a>.</figcaption></figure><p>The figure above shows a clear imbalance across categories. Models perform strongly on relatively boards and widely documented domains such as festivals. In contrast, performance drops sharply on more localized and technical domains.</p><p>Agriculture, Geography, Kinship &amp; Social Hierarchy, and Proverbs/Idioms remain challenging, with most models failing to surpass 50% accuracy, and Agriculture consistently performing worst.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vYAe8S52Pf43v7G4mmjqtg.png" /><figcaption>Performance Comparison: English vs. Nepali vs. Romanized. By <a href="http://raunakregmi.medium.com/">Author</a>.</figcaption></figure><p>The figure shows that Devanagari is often the strongest representation, and for models such as Gemini 3 Flash it can even outperform English. Romanized Nepali is consistently weakest across models, indicating that non-standard transliteration can meaningfully hinder access to culturally grounded knowledge.</p><h3>Does this matter?</h3><p>This isn’t just <strong>academic</strong>. It has real-world consequences.</p><p>With growing scope of LLMs, students using it for <strong>education</strong> must get a proper, regular and culturally accurate answers.</p><p>Nepali speaking users may receive advices that’s missing the local tone and are certainly misleading.</p><p>AI analysis might lead to misinterpretation of local context based on the pre-trained “biased” data used. The dataset the models are currently trained on are Western biased.</p><p>It’s just not about speaking, translating words. It about understanding people’s feeling, understanding the tone, the culture.</p><h3>Conclusion: From Speaking to Feeling</h3><p>AI can speak Nepali, most certainly accurate than most people. It can answer politely.</p><p>But,</p><p>Can it feel the culture? the tone? the meaning behind each words? the tradition?</p><p>Sadly, for now the answer remains a big <strong>NO.</strong></p><p>If Nepal deserves AI that speaks in their mother-tongue, why don’t we deserve one that understands the tradition, respects the hierarchy, knows the cultural etiquette and values and daily realities? Thats the difference between <strong>WORDS</strong> and <strong>FEELINGS</strong>.</p><p>Until then, every casual and confident “हे, के छ?” [“Hey, what’s up?”] to your uncle is a reminder that</p><p><strong>AI speaks — but doesn’t feel.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=755d0db4e428" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/ai-claims-to-know-nepali-culture-or-is-it-just-guessing-755d0db4e428">AI Claims to Know Nepali Culture — Or Is It Just Guessing?</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Know If Your RAG Actually Works]]></title>
            <link>https://ai.plainenglish.io/how-to-know-if-your-rag-actually-works-c75134016cac?source=rss-4f7637b426d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/c75134016cac</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[Raunak Regmi]]></dc:creator>
            <pubDate>Sat, 24 Jan 2026 11:07:49 GMT</pubDate>
            <atom:updated>2026-01-29T19:42:27.094Z</atom:updated>
            <content:encoded><![CDATA[<h4>A practical guide to evaluating retrieval quality, grounding, and performance in real-world RAG systems</h4><h3>Introduction: Why Evaluating RAG Is Harder Than It Looks</h3><p>Retrieval Augmented Generation (RAG) is implemented and used globally to remove the limitations such as static knowledge, hallucinations and lack of specific domain context which large language model (LLMs) often have.</p><p>Integrating external data sources during the generation process, these system promise relatively <strong>more accurate, verifiable, and up-to-date outputs</strong>.</p><p>However, simply implementing a RAG pipeline does not guarantee better results.</p><p>Many teams deploy RAG, often seeing the same results:</p><ul><li>Hallucination still occurs</li><li>Irrelevant documents retrieval</li><li>Inconsistent gains on performance</li></ul><p>This raises a critical question:</p><p><strong>How do you know if your RAG system is actually working?</strong></p><p>This article gives an insight on <strong>technical evaluation focused framework </strong>to answer that question — grounded in established RAG architecture and limitations.</p><h3>What Is Retrieval Augmented Generation (RAG)?</h3><p>Retrieval Augmented Generation is an architecture that combines:</p><ul><li><strong>Information retrieval</strong> from integrated external knowledge sources</li><li><strong>Text generation</strong> with the help of LLM</li></ul><p>Rather than relying just on the parameters the model has trained on, RAG systems use</p><p><strong>NOT STATIC </strong>(like LLMs) but <strong>DYNAMIC</strong></p><p>knowledge-base to retrieve relevant documents and condition the model’s response on that retrieved data</p><p>Let’s analyze it this way,</p><p>Theres an <strong>open-book exam</strong>:</p><ul><li>The system searches the book based on the question</li><li>The retriever selects the most relevant pages from the book</li><li>The LLM combines an answer using selected pages</li></ul><h3>Typical Data Sources</h3><ul><li>PDFs &amp; technical documents</li><li>Internal databases</li><li>Websites and APIs</li><li>Knowledge bases</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1017/1*8cw2jl684UIeme1L00fbMQ.png" /><figcaption>Visual showing LLM vs RAG: Created by <a href="https://www.canva.com/design/DAG6v47ymFk/_SVD-miowkKoSXz6C3MtOQ/edit?utm_content=DAG6v47ymFk&amp;utm_campaign=designshare&amp;utm_medium=link2&amp;utm_source=sharebutton">Author</a></figcaption></figure><ul><li>LLM-only → guessing from memory</li><li>RAG → retrieving documents + generating grounded answers</li></ul><h3>RAG System Architecture (Technical Overview)</h3><p>A standard RAG pipeline consists of four core components:</p><ol><li><strong>Document Store / Vector Database:</strong><br> Stores embedded documents, dividing it into chunks<br> <strong>Vector databases: </strong><em>(FAISS, Pinecone, Chroma, Elasticsearch)</em></li><li><strong>Embedding Model:</strong><br>Converts text → concentrated vector representations</li><li><strong>Retriever:</strong><br> Identifies relevant chunks performing similarity search</li><li><strong>Large Language Models (LLMs)</strong><br> Generates the final answer using the retrieved context</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ghWF32upV9HETKHBWmN9ZQ.png" /><figcaption>End-to-End RAG Pipeline: Created using <a href="https://gemini.google/overview/image-generation/">Gemini 3</a> by Author</figcaption></figure><h3>Key Technical Indicators That Your RAG Works</h3><h3>1. Retrieval Accuracy (The Foundation)</h3><p>If the retrieval fails, the answer generated is totally irrelevant.</p><p>An important principle of RAG evaluation is:</p><p><strong>Wrong documents retrieved = wrong answer</strong></p><p>Well, if so.</p><h4>What to Measure?</h4><ul><li>Do the most relevant documents show up in top-k results?</li><li>Are the retrieved chunks and the query’s topic closely related?</li><li>Are any documents left out?</li></ul><h4>Common Retrieval Metrics</h4><ul><li><strong>Recall@K</strong> — Did we find the relevant documents?</li><li><strong>Precision@K</strong> — How clean is our result?</li><li><strong>MRR (Mean Reciprocal Rank)</strong> — How early did we find the relevant document?</li></ul><p>Poor RAG performance are almost seen due to low retrieval quality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q4xu6zNl8c-n9alnYaZbzg.png" /><figcaption>Precision and recall at K in ranking and recommendations. Image from <a href="https://www.evidentlyai.com/ranking-metrics/precision-recall-at-k">Evidently AI</a></figcaption></figure><h3>2. Context Utilization &amp; Faithfulness</h3><p>If you need your RAG system to work, it <strong>MUST </strong>use the retrieved context.</p><ul><li>The responses are totally based on the retrieved text</li><li>Each claims are linked to a specific chunk</li><li>The model avoids introducing facts that aren’t supported by retreived evidence</li></ul><p>If the retrieved context are ignored by LLMs and answered on basis of the previous knowledge, theres no value added by retrieval.</p><h4>Evaluation Methods</h4><ul><li>Manually matching sources to generated answers</li><li>Evaluating faithfulness using LLMs</li><li>Verifying responses using citations</li></ul><h3>3. Hallucination Reduction Compared to LLM-Only Baselines</h3><p>RAG reduces hallucinations measurably, but not entirely remove them.</p><p>To know so, tests can be done. Such as:</p><p>Run the same queries through:</p><ul><li>LLM only system</li><li>RAG system</li></ul><p>If you see identical response and no differences, the retrieval may not be affecting the generation.</p><p>Note the fact that, effective RAG systems show:</p><ul><li>More precise and concise answers than LLMs</li><li>Fewer made up details (hallucinations)</li><li>Improved accuracy and reliability</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/751/1*k2Ne5EMch3-4toJO1LU8SA.png" /><figcaption>Comparison of RAG to Internal Knowledge. Image from <a href="https://www.pinecone.io/blog/rag-study/">Pinecone</a></figcaption></figure><h3>4. Sensitivity to Document Quality</h3><p>A reliable RAG system should have predictable performance gains as quality of data improves.</p><p>Test whether:</p><ul><li>Well document formatting leads to better answers</li><li>Improved chunking provides clearer responses</li><li>Metadata increases relevant retrieval</li></ul><p>If improving documents do not change the results, the issue might lie with the retriever or embedding quality.</p><h3>Common Technical Pitfalls in RAG Systems</h3><p>Think your RAG pipeline is well-designed?</p><p>But common issues may fail it, such as:</p><h4>Poor Chunking Strategy</h4><ul><li>Large chunks → more noise</li><li>Small chunks → lost context</li></ul><h4>Low-Quality Source Documents</h4><ul><li>Scanned PDFs</li><li>Bad document formatting</li><li>Missing well-defined structure</li></ul><h4>Overestimating RAG Capabilities</h4><ul><li>RAG is <strong>NOT</strong> a replacement for fine-tuning (RAG ≠ fine-tuning)</li><li>Bad data can <strong>NOT </strong>be fixed by RAG (Bad data in → bad answers out)</li><li>LLM’s reasoning quality affects RAG.</li></ul><h3>Retrieval Evaluation</h3><ul><li>Analyzing vector similarity score distributions</li><li>Testing relevance between document and query</li><li>Offline benchmark query evaluation</li></ul><h3>Generation Evaluation</h3><ul><li>Human evaluation using scoring criteria</li><li>LLM used as a judge</li><li>Faithfulness and relevance metrics</li></ul><h3>System-Level Metrics</h3><ul><li>Total response time, request → output</li><li>Cost per query</li><li>Rate of failed retrieval</li></ul><p>Evaluating all three layers makes sure that the RAG system is ready for production and for the industry.</p><h3>Conclusion: From RAG Demo to Production System</h3><p>A Retrieval Augmented Generation system works only if the <strong>retrieval, grounding and generation</strong> work together.</p><p>Want to say your RAG works?</p><p>Well, for that, your system must:</p><ul><li>The generated answers must link to the source documents</li><li>Evaluate accuracy of the retrieved results</li><li>Prove reduced hallucinations</li><li>Give improved answers, on improvement of data quality</li></ul><p>RAG isn’t some magic it is just <strong>engineering applied to an AI aware about its knowledge</strong>.</p><p>When implemented properly, it becomes one of the most powerful tool. The most reliable architectures for the real-world system.</p><p>Also for further knowledge, check out my presentation on <a href="https://www.canva.com/design/DAG6v47ymFk/_SVD-miowkKoSXz6C3MtOQ/edit?utm_content=DAG6v47ymFk&amp;utm_campaign=designshare&amp;utm_medium=link2&amp;utm_source=sharebutton">RAG</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c75134016cac" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/how-to-know-if-your-rag-actually-works-c75134016cac">How to Know If Your RAG Actually Works</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Businesses Are Moving to the Cloud: An AWS Perspective]]></title>
            <link>https://aws.plainenglish.io/why-businesses-are-moving-to-the-cloud-an-aws-perspective-7d96e654cb92?source=rss-4f7637b426d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d96e654cb92</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[amazon-web-services]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <dc:creator><![CDATA[Raunak Regmi]]></dc:creator>
            <pubDate>Thu, 15 Jan 2026 15:00:31 GMT</pubDate>
            <atom:updated>2026-01-15T18:39:42.383Z</atom:updated>
            <content:encoded><![CDATA[<p>Understanding cloud computing fundamentals and why Amazon Web Services is transforming how we build digital solutions</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sgklRPMY1qCwvtTSs1TP4Q.png" /><figcaption>Industries Moving to Cloud. Created using <a href="https://gemini.google/overview/image-generation/">Gemini 3</a> by Author.</figcaption></figure><p>Today’s world, businesses are growing rapidly, scaling quickly, and protecting data more securely and efficiently with the help of cloud storage. From being a useful solution to essential needs, cloud has changed the perspective of every business owner. Transitioning from the hardened, traditional method of scaling the servers, costly and irrelevant data protection.</p><p>Among all the cloud services, undoubtedly <strong>AWS (Amazon Web Services)</strong> has been the growing leader of this century. From start-up to multinational businesses everybody is using it.</p><p>So why are companies changing in this way?</p><p>Let’s dissect it from a practical, real-world standpoint.</p><h3>Understanding Cloud Computing (Without Jargon)</h3><p>Before even getting into some fancy jargons and words, let’s dive shortly into what “the cloud” means.</p><p>At its core, cloud refers to accessing physical hardwares such as — databases, servers, networks, softwares over the internet by renting them instead of owning them.</p><p>Instead of making significant investments in data centers, companies</p><ul><li>Rent only what they need</li><li>Scale resources up or down instantly</li><li>Pay for what they use</li></ul><p>AWS offering these services on-demand and inexpensive, makes the businesses <strong>work on innovation</strong> and creativity instead of focusing on <strong>infrastructures</strong>.</p><h3>Cost Efficiency: Pay for What You Use</h3><p>One of the main <strong>shifting</strong> forces towards cloud adoption is cost efficiency and optimization.</p><p>Standard IT infrastructure needs:</p><ul><li>High initial hardware expenses</li><li>Continuous scaling and updates</li><li>Highly provisioned to meet needs and demands</li></ul><p>Businesses switch to a <strong>pay-as-you-go</strong> model with AWS. This implies:</p><ul><li>No initial and fixed capital requirement</li><li>Highly reduced operational expenses</li><li>Increased predictability of the budget</li></ul><p>This model helps small start-ups and financial organizations remove financial barriers and keep-up with the expenditure. Later, allowing them to improve cost-efficiency &amp; <strong>resource allocation.</strong></p><h3>Scalability and Flexibility at Any Speed</h3><p>Almost never are the <strong>business requirements</strong> ever the same. Growth of a company, peak inflation, unexpected expenditures can often be scary.</p><p>This is resolved by AWS’s on-demand scalability:</p><ul><li>Servers should be scaled in minutes rather than months.</li><li>Manage <strong>traffic spikes</strong> automatically</li><li>Grow internationally without constructing new infrastructure</li></ul><p>This is why everybody <strong>loves</strong> AWS.</p><p>AWS ensures you that the system can change in minutes rather than it taking months, also with the growth and use of users accordingly and accurately.</p><p><strong>The best part,</strong> it doesn&#39;t matter whether it’s a mobile app, digital marketing campaign, or even a product launch.</p><h3>Security and Compliance You Can Trust</h3><p>Often, security is the biggest issue that comes up while companies are moving to the cloud. Surprisingly, AWS smoothly deals with it.</p><p>AWS provides:</p><ul><li>Integrated encryption</li><li>Management of identity and access <strong>(IAM)</strong></li><li>Threat detection and network firewalls</li><li>Sticking to international standards (ISO, GDPR, HIPAA, and more)</li></ul><p>Because security is AWS’s full-time responsibility, many businesses actually find that AWS offers <strong>stronger security</strong> than their traditional setups.</p><h3>Faster Innovation and Time to Market</h3><p>Surely, being a big business, speed is necessary. Businesses that innovate more quickly have a competitive advantage.</p><p>AWS enables rapid innovation by offering:</p><ul><li>Fully managed databases</li><li>Serverless computing with AWS Lambda</li><li>AI and machine learning services</li><li>DevOps and CI/CD tools</li></ul><p>Having to wait for the infrastructural setup or approval is the worst. But, developers can build, test and deploy without wasting anytime waiting, reducing time to market.</p><h3>Reliability and Global Reach</h3><p>AWS runs data centers in several different parts of the world. Thus, companies gain from:</p><ul><li>High accessibility</li><li>Options for disaster recovery</li><li>Low-latency for consumers worldwide</li></ul><p>AWS enables worldwide deployment easy and dependable for businesses growing globally.</p><figure><img alt="Reliability and Global Reach. Created using Gemini 3 by Author." src="https://cdn-images-1.medium.com/max/1024/1*Zu9_jsa5biXW5VVE5M06mw.png" /><figcaption>Reliability and Global Reach. Created using <a href="https://gemini.google/overview/image-generation/">Gemini 3</a> by Author.</figcaption></figure><h3>Real-World Use Cases of AWS Cloud Adoption</h3><p>Businesses across industries are moving to AWS:</p><ul><li><strong>Startups</strong> use AWS to build scalable products without upfront costs</li><li><strong>E-commerce platforms</strong> handle peak traffic effortlessly</li><li><strong>Enterprises</strong> modernize legacy systems</li><li><strong>Media and gaming companies</strong> deliver content globally with low latency</li></ul><p>AWS isn’t just for tech companies, it’s for any business that relies on digital systems.</p><h3>The Future Is Cloud-First</h3><p>Cloud computing is no longer the future, it is the present.</p><p>With AWS, it has made this successful. As the digital world is transforming rapidly, it has been that businesses which delay in adopting cloud technology are certainly falling behind in innovation, agility and security.</p><p>From scalability and cost efficiency to reliability and speed, AWS provides a strong foundation for modern businesses competing in a digital-first world.</p><h3>Final Thoughts</h3><p>Cloud migration is not just a decision, it is a business strategy.</p><p>The cloud, and AWS in particular, are essential for a company to be in the digital market delivering better experiences among the customers, improving products and for sustainable growth.</p><p>Cloud is an essential for business aiming to stay at the competitive edge with flexibility and ready for the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d96e654cb92" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/why-businesses-are-moving-to-the-cloud-an-aws-perspective-7d96e654cb92">Why Businesses Are Moving to the Cloud: An AWS Perspective</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>