<?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 Jesus Almaraz Martin on Medium]]></title>
        <description><![CDATA[Stories by Jesus Almaraz Martin on Medium]]></description>
        <link>https://medium.com/@jesusalmarazmartin?source=rss-4a2b13c05138------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*YAKYV894_qy82UZ7lJlQew.png</url>
            <title>Stories by Jesus Almaraz Martin on Medium</title>
            <link>https://medium.com/@jesusalmarazmartin?source=rss-4a2b13c05138------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 23:46:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jesusalmarazmartin/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[How to make useful Unified Microsoft Copilot Experience in Business Central]]></title>
            <link>https://jesusalmarazmartin.medium.com/how-to-make-useful-unified-microsoft-copilot-experience-in-business-central-e460bb9dbe21?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/e460bb9dbe21</guid>
            <category><![CDATA[co-pilot]]></category>
            <category><![CDATA[als]]></category>
            <category><![CDATA[business-central]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 17:49:22 GMT</pubDate>
            <atom:updated>2026-11-08T17:49:22.144Z</atom:updated>
            <content:encoded><![CDATA[<h3>In-Context AI Assistant</h3><p>I am absolutely dependent on all tools I work with, in having an In-Context AI Assistant. If you use some apps like Gitlab, you already know what is an In-Context AI Assistant: An AI assistant specific for the web app you are using. This kind of assistant have 3 basis:</p><ul><li>UI and page context retrieval.</li><li>Use of specific MCP and tools.</li><li>LLM chat connections.</li></ul><p>This way when you query the assistant the “How to”, he tries to answer by scrapping your web page context and consulting tools, to get a good response based only in the context of the app we are using.</p><h3>Unified Microsoft Copilot Experience</h3><p>From incoming Business Central 29 release the Assistant tool (is) will be Unified Microsoft Copilot Experience:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eD6BDcM7YmEP8EjUTmBMRw.png" /></figure><p>Two important points of this assistant:</p><ul><li>Do not act with BC. Is not an agent, only a tool for Data and Procedural Knowledge Retrieval. Is read only.</li><li>I try to check this very carefully, but it doesn’t hurt if you double check this affirmation: the token consumption is free and included in the BC license.</li></ul><p>For data retrieval (Get customer data, Item data, sales statics, etc.) it has access to exposed API and MCP.</p><p>But in this post I want to focus on Procedural Knowledge Retrieval.</p><h3>Customizations Procedural Knowledge Retrieval</h3><p>This is the “How to…” question. The user asks the chat user how to fix an error, set a value, make an operation (in this moment not to do it, remember not agentic features here), app capabilities.</p><p>From developers view, I think is more interesting the question how assistant retrieves the information, how we can help the assistant to discover our customized features.</p><p>Investigating I got the answer that we have not hooks or events subscriptions to adapt the chat response (but I will keep searching for this kind of hack). So in the documentation I found only two ways to inject this knowledge into the assistant:</p><ul><li>Tooltips.</li><li>Help Url.</li></ul><p>Reminder:</p><p>A tooltip are extended instructions for a page a text control that hover when you are placed in:</p><pre>field(URL; Rec.URL)<br>{<br>  ToolTip = &#39;Specifies the value of the URL field.&#39;;<br>}</pre><p>I left this terrible Tooltip to show an stupid Tooltip. This tooltip says absolutely nothing, and is the default tooltip some automatizations create.</p><p>The context help is this URLs settled in app.json:</p><pre>  &quot;help&quot;: &quot;&quot;,<br>  &quot;url&quot;: &quot;&quot;</pre><h3>Conclusion</h3><p>If you want useful interactions with the assistant for your AL customizations write good Tooltips and good help URLs.</p><p>Additional work? Very little additional work. Long time ago I adopted the attitude of making things for AI with AI. Make a skill/agent in Vscode for doing these tooltips and Help, from App workspace and I think the result will be great and the cost minimal. Do not forget to translate to your language after.</p><p>I Don’t promise it (I have lots of posts pending) but I would write a hand off about tools to build easy tooltips and help.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e460bb9dbe21" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Style Guide: Gemini Integration with Business Central]]></title>
            <link>https://jesusalmarazmartin.medium.com/style-guide-gemini-integration-with-business-central-b38bc7138edb?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/b38bc7138edb</guid>
            <category><![CDATA[gemini-api]]></category>
            <category><![CDATA[business-central]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 04:22:42 GMT</pubDate>
            <atom:updated>2026-09-25T04:22:42.776Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>In this post we will reach these goals…or try it:</p><ul><li>Bring a real use case for AI in Business Central.</li><li>In my previous posts I always used Gemini API integration without SDK. This way we can connect with Gemini and include AI capabilities from any system.</li><li>Use the nowadays cheaper AI provider, Gemini, to reach these features.</li></ul><h3>The use case</h3><p>The description of the General ledger transactions must follow an standardized guide to comply with company and tax authorities demands. This matches well with AI capabilities: the question to the machine is what the next word is, if LLM does not fit with this task I do not understand anything.</p><p>So we have to set a style guide in the system for journal descriptions. When we validate the description the system will call Gemini asking if the description follows the guide, and the answer includes a score 1 to 10, the failures of your description, and an improved description. If the score does not reach a limit the system will:</p><ul><li>Shows us a message with the score, the alternative and the score explanation.</li><li>If accepted, the new and improved description substitutes the previous description.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D79hacmEdZLiESZ8mexiDg.gif" /></figure><h3>Use and Configuration</h3><p>For configuration we must go to Search/Style Guides, and fill the fields:</p><ul><li>Table No.</li><li>Field No.</li><li>URL of the Gemini service: <a href="https://generativelanguage.googleapis.com/v1beta/models/gemini-3.5-flash-lite:generateContent">https://generativelanguage.googleapis.com/v1beta/models/gemini-3.5-flash-lite:generateContent</a></li><li>API Key: gdfashjfs67884hjllsie9. You can use this key and if it works, please notify me.</li><li>System Prompt: you know what this works, you set here a role, the rules that text must follow, and with further tests improve it.</li></ul><h3>Business Central implementation</h3><p>All the project is public: <a href="https://github.com/JalmarazMartn/BC-Gemini">JalmarazMartn/BC-Gemini</a></p><p>There is no lots of code so I want to highlight the implementation keys:</p><p>As we can see in all my posts, the API connection is done without SDK, this way we can use it in all systems with a simple HttpRequest. In BC is so simple as convert JavaScript code to AL Business Central Code, changing fetch for this code:</p><pre>URLFinal := URL + &#39;?key=&#39; + ApiKey;<br><br>// Prepare the HTTP body content and set Content-Type header<br>HttpContent.WriteFrom(RequestBodyText);<br>HttpContent.GetHeaders(HttpHeaders);<br>HttpHeaders.Remove(&#39;Content-Type&#39;);<br>HttpHeaders.Add(&#39;Content-Type&#39;, &#39;application/json&#39;);<br><br>// Make the POST request<br>if HttpClient.Post(URLFinal, HttpContent, HttpResponseMessage) then</pre><p>The second key is and event subscription to General Journal Line Description Validate:</p><pre>    [EventSubscriber(ObjectType::Table, Database::&quot;Gen. Journal Line&quot;, &#39;OnAfterValidateEvent&#39;, &#39;Description&#39;, false, false)]<br>    local procedure CheckDescription(var Rec: Record &quot;Gen. Journal Line&quot;)<br>    begin<br>        if not StyleGuide.get(Rec.RecordId.TableNo, Rec.FieldNo(Description)) then<br>            exit;<br>        RequestBodyText := BuildStyleGuideRequestBody(rec.Description, StyleGuide.GetSystemPrompt());<br>        ResponseText := PostToGemini(StyleGuide.&quot;API Key&quot;, RequestBodyText, StyleGuide.URL);<br>        ParseStyleResponse(ResponseText, CumpleGuia, Sugerencia, MayorError);<br>        if CumpleGuia = 10 then<br>            exit;<br>        if not Confirm(SubsTextQst, true, CumpleGuia, MayorError, Sugerencia) then<br>            exit;<br>        rec.Description := CopyStr(Sugerencia, 1, MaxStrLen(rec.Description));<br>    end;</pre><p>The rest of the work is to parse and create JSON objects, for body and headers. As we do in <a href="https://medium.com/@jesusalmarazmartin/how-to-automate-style-guide-compliance-part-2-improving-gemini-api-calls-491b222b6570">How to Automate Style Guide Compliance part 2: improving Gemini API calls | by Jesus Almaraz Martin | Sep, 2026 | Medium</a> we ask for a formatted output for get a less probabilistic response.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b38bc7138edb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Automate Style Guide Compliance part 2: improving Gemini API calls]]></title>
            <link>https://jesusalmarazmartin.medium.com/how-to-automate-style-guide-compliance-part-2-improving-gemini-api-calls-491b222b6570?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/491b222b6570</guid>
            <category><![CDATA[gemini-api]]></category>
            <category><![CDATA[gemini]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 16:18:54 GMT</pubDate>
            <atom:updated>2026-09-06T16:18:54.350Z</atom:updated>
            <content:encoded><![CDATA[<p>Almost a year ago I published <a href="https://medium.com/@jesusalmarazmartin/how-to-automate-style-guide-compliance-cca25b1a7aa9">How to Automate Style Guide Compliance | by Jesus Almaraz Martin | Medium</a> . Let’s do a little remainder: was all about checking if a text follows a style guide, using API Gemini call. The full description of this stuff is in the post, you can read it in a few minutes. I am too lazy to write long texts.</p><p>The subject of this new post is to use some AI engineering concepts to improve the system:</p><ul><li>Let a context to set a state to the queries.</li><li>Set the output structure.</li><li>Reduce latency.</li></ul><p>Note that in my set of improvements and engineering is not included to improve the prompt. The prompt was decent but not very good. If you use the same prompt you will feel the non deterministic nature of this animal, but this is not the target of this post. You know the tip to improve every prompt: do it with AI.</p><h3>Setting the context: gimme you credit card</h3><p>When you use the Gemini API there is no state saving. So it could be good to save the system prompt between calls, for speed up (not much) and save tokens (very much if you believe the official documentation). Only two warnings:</p><ul><li>You must enter your pay form in google, a credit card.</li><li>There is a minimum of tokens for setting the context. If your context is too small you can set it.</li></ul><p>The API call to set the context is:</p><pre>                const response = await fetch(&#39;https://generativelanguage.googleapis.com/v1beta/cachedContents?key=&#39; + apiKey, {<br>                    method: &#39;POST&#39;,<br>                    headers: {<br>                        &#39;Content-Type&#39;: &#39;application/json&#39;<br>                    },<br>                    body: JSON.stringify({<br>                        model: &#39;models/gemini-3.5-flash-lite&#39;,<br>                        displayName: &#39;style_guide_system_instruction&#39;,<br>                        ttl: &#39;86400s&#39;,<br>                        systemInstruction: {<br>                            parts: [{<br>                                text: systemInstructionText//All the style guide<br>                            }]<br>                        }<br>                    })<br>                });</pre><p>The response gives us a context ID. When we query API next time we send it this way:</p><pre>                if (cachedContentId) {<br>                    requestBody.cachedContent = cachedContentId;<br>                } else if (includeInlineSystem) {<br>                    requestBody.systemInstruction = { parts: [{ text: systemInstructionText }] };<br>                }</pre><h3>Set the output format</h3><p>We can set an output format for the response. It is obvious this will improve the response an ease to handle it:</p><pre>                const requestBody = {<br>                    contents: [{<br>                        role: &#39;user&#39;,<br>                        parts: [{ text: prompt }]<br>                    }],<br>                    generationConfig: {<br>                        responseMimeType: &quot;application/json&quot;,<br>                        responseSchema: {<br>                            type: &quot;OBJECT&quot;,<br>                            properties: {<br>                                CumpleGuia: {<br>                                    type: &quot;INTEGER&quot;,<br>                                    description: &quot;Puntuación de 1 a 10&quot;<br>                                },<br>                                Sugerencia: {<br>                                    type: &quot;STRING&quot;,<br>                                    description: &quot;como debería haberse redactado mejor el texto: texto altenativo sin más explicaciones&quot;<br>                                },<br>                                MayorError: {<br>                                    type: &quot;STRING&quot;,<br>                                    description: &quot;Error principal&quot;<br>                                }<br>                            },<br>                            required: [&quot;CumpleGuia&quot;, &quot;Sugerencia&quot;, &quot;MayorError&quot;]<br>                        }<br>                    }</pre><h3>Reducing latency</h3><p>My first calls were too slow for my goal: introduce this system in ERP, when editing text in records.</p><p>But the answer for this point is easy: use the Flash-Lite model. We do not need the smartest guy in the bar for my purpose, so Lite is enough intelligent for this task:</p><pre>fetch(&#39;https://generativelanguage.googleapis.com/v1beta/models/gemini-3.5-flash-lite</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=491b222b6570" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Read the book “Building AI-Powered Apps with Angular”.. but not for Angular]]></title>
            <link>https://jesusalmarazmartin.medium.com/read-the-book-building-ai-powered-apps-with-angular-but-not-for-angular-35b042f7b8aa?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/35b042f7b8aa</guid>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[gemini-api]]></category>
            <category><![CDATA[web-development]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 09:10:30 GMT</pubDate>
            <atom:updated>2026-08-30T09:10:30.255Z</atom:updated>
            <content:encoded><![CDATA[<p>Read the book “Building AI-Powered Apps with Angular”.. but not for Angular</p><h3>Introduction</h3><p>I cannot explain why I read the book “Building AI-Powered Apps with Angular” by authors Giorgio Boa and Fabio Biondi. Not very interested in Angular or Frontend developing. I leave this stuff to AI (just kidding, but not much).</p><p>But I read the preview and then I bought it. This is the reason is a very interesting book: not a book about angular, is a great and brief introduction to AI engineering, and how to include Gemini in your apps.</p><p>Because for focusing in many times to AI Engineering, and the way the examples worked, this book could stand the test of time.</p><p>LLMs are great at two things: losing money and making you feel like you are missing something. So I want books that can give a deeper understanding.</p><h3>Book Highlights</h3><h4>AI Engineering content</h4><ul><li>Not the target of the book, but starts with a great introduction about Machine learning basics: Supervised, Unsupervised, and Reinforcement learning. Neural networks concepts too.</li><li>Note that right now almost every time we say AI we mean LLM. But there is AI beyond LLM. LLM is the current approach to use the AI, but AI is more than LLM. I never get tired of repeat this, but I think is very important. After the introduction, the book focuses in LLM.</li><li>So the book introduces concepts about LLMs: Tokenization, Top-K, Top-P, temperature and a great introduction to prompt engineering.</li></ul><h4>Gemini API</h4><p>Note: This book mainly use the SDK instead of fetch or Http Request to access the Gemini API: I do not like it, because I am a Business Central Developer, and I need always to use less libraries and resources. Not a problem, because this change from SDK to standard request is very easy to do with AI.</p><p>The book develops web apps integrating AI capabilities calling Gemini API. Is plenty of very useful examples and code, but for me the main knowledge I got, was about setting the parameters before calling to optimize results and save tokens:</p><ul><li>You can pass Core Generation parameters: temperature, tokens.</li><li>“ResponseMimeType”, for setting structured output.</li></ul><pre>&quot;generationConfig&quot;: {<br>    &quot;temperature&quot;: 0.1,<br>    &quot;topP&quot;: 0.95,<br>    &quot;maxOutputTokens&quot;: 1024,<br>    &quot;responseMimeType&quot;: &quot;application/json&quot;,<br>    &quot;responseSchema&quot;: {<br>      &quot;type&quot;: &quot;OBJECT&quot;,<br>      &quot;properties&quot;: {<br>        &quot;title&quot;: { &quot;type&quot;: &quot;STRING&quot; },<br>        &quot;specs&quot;: { <br>          &quot;type&quot;: &quot;ARRAY&quot;, <br>          &quot;items&quot;: { &quot;type&quot;: &quot;STRING&quot; } <br>        }<br>      },<br>      &quot;required&quot;: [&quot;title&quot;, &quot;specs&quot;]<br>    }<br>  }<br>}</pre><ul><li>Other interesting thing is to force the use of links included in the prompt to get updated information with “urlContext”. Another way to do it with less token consumption is setting google search as tool.</li></ul><pre>&quot;tools&quot;: [<br>    {<br>      &quot;googleSearch&quot;: {}<br>    }</pre><h4>Extra bonus</h4><p>In addition two great chapters caught my attention:</p><ul><li>MCP: I always thought MCP was for serving tools to AI agents. But this is not true, the book teaches us that MCP can serve 4 capabilities: Tools, Resources, Prompts and Sampling.</li><li>RAG: great example about how to make a RAG system from scratch.</li></ul><h3>Conclusion</h3><p>A worth buying: you can get a fast tour with it of AI engineering basics, and learn how to introduce AI tools in your apps using Gemini API calls.</p><h3>Links</h3><p>Of course the link of the book:</p><p><a href="https://www.oreilly.com/library/view/building-ai-powered-apps/9781806383498/">Building AI-Powered Apps with Angular [Book]</a></p><p>If you want to dive deeper in AI engineering the most popular book of the subject from Chip Huyen:</p><p><a href="https://www.oreilly.com/library/view/ai-engineering/9781098166298/">AI Engineering [Book]</a></p><p>My own introduction to Machine Learning:</p><p><a href="https://medium.com/analytics-vidhya/machine-learning-for-programmers-f840e463886e">How can programmers understand Machine Learning paradigm | by Jesus Almaraz Martin | Analytics Vidhya | Medium</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=35b042f7b8aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to use local LLM in VSCode: Feeling the pain]]></title>
            <link>https://jesusalmarazmartin.medium.com/how-to-use-local-llm-in-vscode-feeling-the-pain-5ea2ebdc9a28?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ea2ebdc9a28</guid>
            <category><![CDATA[ollama]]></category>
            <category><![CDATA[vscode-extension]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 21 Jun 2026 09:07:28 GMT</pubDate>
            <atom:updated>2026-06-21T09:07:28.628Z</atom:updated>
            <content:encoded><![CDATA[<p>Disclaimers:</p><ul><li>First part of the post is mostly opinion.</li><li>Second and tech part of the post is accurate, but considere that the test are done in a 9 nine years PC, with the minimal features. I only need navigate and VSCode. I am an adult, do not need gamming.</li></ul><h3>Do not say AI, say LLM</h3><p>Years ago AI was a promising thing, with lots of great applications: speech to text, image recognition and OCR, assistants, predictive models. Was the era of machine learning, AlphaFold and greats things that few people knew and used.</p><p>Then LLM comes. Someone though that would be a great idea to do an AI tool for all. The same tool that help you in code development, can make a poem, give you psychologic advice and film recommendations.</p><p>In these confusing years I have the feeling that the main thing LLM do better is losing money. I want to explain my self, I love LLM, I use LLM in my work, but there are problems that make the LLMs an unreliable partner (IMO):</p><ul><li>The non-deterministic nature of the animal. This is discussed in <a href="https://medium.com/@jesusalmarazmartin/github-copilot-doing-fine-and-not-so-fine-2-f63e594f5f12">GitHub Copilot: doing fine and not so fine (2) | by Jesus Almaraz Martin | Medium</a></li><li>The speed. Completion code runs ok, but other chat is too slow for me.</li><li>The cost has not been already transferred and we do not know how much will be the bill.</li></ul><h3>The pain in my PC: local model</h3><p>LLMs have a terrible cost of energy and processing, and nowadays for the final user is free or like Github-Copilot, you only pay for a flat-rate subscription. This cannot be eternal.</p><p>I think that you do not need the smartest model to help you to program, so want to know if I can use free local model, discover how to set it and test its performance.</p><p>I Installed the Ollama model quwen2.5-coder7b from PowerShell:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/726/1*WY4fK1YwH8h3mhBgKoeK2g.png" /></figure><p>And then I set this model in Vscode:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/340/1*pF_FzJ2gaBkkz0h68288hw.png" /></figure><p>After push setup symbol above we set new model in chat:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/353/1*7DEibrOUwaYdi-PsRLVg_Q.png" /></figure><p>We are ready to use it. So if we type in VSCode chat this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/456/1*s-xZhlhMz7CVwzl8YspUWA.png" /></figure><p>We begin to feel the pain, as I said my desktop is old, but the response for a simple question from chat takes over 15 minutes. And the ugliest part of this, is that my software throw the towel for doing anything else during the process, becoming absolute collapsed.</p><p>If I ask the same in the Ollama native chat i get the response, not fast, but fastest. The quality of the responses are not like On-line big models, but for JavaScript coding are good enough to help us:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/867/1*GnLtUq8KHD55G8hkM6jrGQ.png" /></figure><h3>Lab: using VSCode API extension</h3><p>In the search of faster performances I tried this test:</p><p>I send the prompt through code using LM VSCode API object in a VSCode extension:</p><pre>        const ollamaUrl = getOllamaUrlConfiguration() || &#39;http://localhost:11434&#39;;<br>        const ollamaModel = getModelFamilyConfiguration() || &#39;qwen2.5-coder:7b&#39;;<br>        <br>        const fullPrompt = <br>            prompt.taskExplanation + &#39;\n&#39; +<br>            &#39;Example: when input:&quot;&#39; + prompt.inputExample + &#39;&quot;, the output response must be:&quot;&#39; + prompt.outputExample+&#39;&quot;&#39; + &#39;\n&#39; +<br>            prompt.inputIndications + &#39; &#39; + inputText;<br>        <br>        const response = await fetch(`${ollamaUrl}/api/generate`, {<br>            method: &#39;POST&#39;,<br>            headers: {<br>                &#39;Content-Type&#39;: &#39;application/json&#39;,<br>            },<br>            body: JSON.stringify({<br>                model: ollamaModel,<br>                prompt: fullPrompt,<br>                stream: false,<br>            })<br>        });</pre><p>With this implementation, the response of the local model is faster, 6 minutes not bad considering the limitations of the environment, so I think this could be the way to get some extra speed: working your own chat panel to work with these models (and a local server or a better PC).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ea2ebdc9a28" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GitHub Copilot: doing fine and not so fine (2)]]></title>
            <link>https://jesusalmarazmartin.medium.com/github-copilot-doing-fine-and-not-so-fine-2-f63e594f5f12?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/f63e594f5f12</guid>
            <category><![CDATA[productivity-hacks]]></category>
            <category><![CDATA[github-copilot]]></category>
            <category><![CDATA[vscode]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Fri, 27 Feb 2026 06:49:44 GMT</pubDate>
            <atom:updated>2026-02-27T06:49:44.772Z</atom:updated>
            <content:encoded><![CDATA[<p>Follow up of <a href="https://medium.com/@jesusalmarazmartin/github-copilot-doing-fine-and-not-so-fine-75ba5db370c8">GitHub Copilot: doing fine and not so fine | by Jesus Almaraz Martin | Nov, 2025 | Medium</a></p><h3>Disclaimer</h3><p>Things improve and improve, so note carefully the date of this post. I will try to keep it updated but can be outdated sometimes, since sometime.</p><p>Another disclaimer: this could seem a fast close of the subject. Because it is a fast close. Months ago the first post, I notice that I can wait, and add more content and use cases, but this will be an eternal wheel. So, let´s close and wait for what the future will bring us.</p><h3>User manual, code explanations and comments</h3><p>In point is no brainer: this is a strength of IA. And the comments come in the more confortable way, in my opinion, in auto-completion mode.</p><h3>Powershell</h3><p>An specific use: I need Powershell scripts, but i am not very skillful making them. But Co-Pilot is very good in this matter and even gave me refracting of the generated code.</p><h3>Generating test data</h3><p>In the example <a href="https://github.com/JalmarazMartn/GeminiAPI/blob/master/genetateTestData.js">GeminiAPI/genetateTestData.js at master · JalmarazMartn/GeminiAPI</a> I ask Gemini API to create Data with this prompt. Notice that I am not giving too much info to the LLM:</p><pre>        &quot;parts&quot;: [<br>          {<br>            &quot;text&quot;: &quot;Generate test data for a customer database <br>            with 10 entries. Each entry should include a <br>            unique customer ID, full name, email address, <br>              phone number, and physical address.&quot; + <br>            &quot;Must be spanish names and addresses from Madrid&quot; +<br>            &quot;Format the output as a JSON array: [{\&quot;CustomerID\&quot;: , <br>            \&quot;FullName\&quot;: , \&quot;Email\&quot;: , \&quot;PhoneNumber\&quot;: , <br>            \&quot;Address\&quot;:  }]&quot;<br>          }</pre><p>And the response:</p><pre>[<br>  {<br>    &quot;CustomerID&quot;: &quot;MAD-001&quot;,<br>    &quot;FullName&quot;: &quot;Alejandro García Fernández&quot;,<br>    &quot;Email&quot;: &quot;a.garcia.fer@example.es&quot;,<br>    &quot;PhoneNumber&quot;: &quot;+34 912 34 56 78&quot;,<br>    &quot;Address&quot;: &quot;Calle de Alcalá, 45, 28014 Madrid&quot;<br>  },<br>  {<br>    &quot;CustomerID&quot;: &quot;MAD-002&quot;,<br>    &quot;FullName&quot;: &quot;Lucía Martín Sánchez&quot;,<br>    &quot;Email&quot;: &quot;lucia.martin88@example.com&quot;,<br>    &quot;PhoneNumber&quot;: &quot;+34 915 67 12 34&quot;,<br>    &quot;Address&quot;: &quot;Paseo de la Castellana, 120, 28046 Madrid&quot;<br>  },<br>  {<br>    &quot;CustomerID&quot;: &quot;MAD-003&quot;,<br></pre><h3>Web Security and performance…if you are not a great web developer</h3><p>In my post <a href="https://medium.com/@jesusalmarazmartin/ocr-browser-extension-how-to-make-a-browser-extension-07232a323b8c">OCR browser extension: How to make a browser extension | by Jesus Almaraz Martin | Feb, 2026 | Medium</a> I made a Browser extension and I made it mainly using AI.</p><p>The last step was asking AI to improve app security and performance and the result was good. First improved performance:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/311/1*LVZExh-6320RGZUgMHMUXA.png" /></figure><p>And then Improve security:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/313/1*v57VHgQQfCmJHfDftK0ZAg.png" /></figure><h3>Code Fixes</h3><p>Fixes enter in the “not so fine zone”. A simple code fix as a forgotten quote, or so on, is OK, but another kind of code fixing, that requieres knowledge and more important, accuracy, is better done, and faster by an expert. Even if you have an agent or a previous made prompt to apply fixes can be complicated and the results must be carefully supervised.</p><h3>Tech stuff vs Business logic</h3><p>In my AL Business Central, I develop lots of Business logic, with implicit account. This is not very reliable, and in this case when I mechanized the code generation.</p><p>The tech stuff is more reliable, as I said above in other sections.</p><h3>Conclusion: Doing that you do not do ok yet, Doing that you already do ok, overfit?</h3><p>Over and over when we talk about AI, we return to the same ideas:</p><ul><li>We are talking about probabilistic tools, no deterministic , so we can not trust in reduce a constant amount of work using AI. The AI reduce the work, but not in an accurate way, and this is a problem in a world of budgets and delivery terms. Some tools, like skills, MCP, LLM Tools or agents could put on the table more accuracy but we must try and wait.</li><li>It feels like code-generation LLMs are overfitted: tasks that look equally complex can result in wildly different levels of quality. I try to do everything with AI, but if do not work I give up quickly.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f63e594f5f12" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OCR browser extension: How to make a browser extension]]></title>
            <link>https://jesusalmarazmartin.medium.com/ocr-browser-extension-how-to-make-a-browser-extension-07232a323b8c?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/07232a323b8c</guid>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[browser-extension]]></category>
            <category><![CDATA[javascript]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 01 Feb 2026 06:35:30 GMT</pubDate>
            <atom:updated>2026-02-01T06:43:05.701Z</atom:updated>
            <content:encoded><![CDATA[<h3>Use case</h3><p>In our support work, users often send test data — order numbers, item codes, and, my least favorite, 18‑digit SSCCs. The problem is they usually share this information as an image, so I end up typing the codes manually just to run a test. For long identifiers like SSCCs, or combinations such as order + customer + item number, it’s much better to have a little help than to key everything by hand. Let see a not real SSCC image example.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/414/1*asw9Nj2t6jilJPK0tGqWAQ.png" /><figcaption>From <a href="https://www.clearmark.uk/resources/advice-and-knowledge/sscc">SSCC for Pallet Labelling — Advice and Knowledge | Clearmark</a></figcaption></figure><p>I want to have a tool in the browser to avoid write this code manually. So, I made a browser extension to:</p><ul><li>Open a side panel.</li><li>When capture screen, or an area of the screen, press side panel button “Clipboard image OCR scan”.</li><li>Then the OCR leaves the scanned text, to copy it from the textbox.</li></ul><p>Let see the extension in action, could work that way in combination of screen clipboard standard tools:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vJlwmFQzMDJAnsf2Tv4TrA.gif" /><figcaption>OCR in action</figcaption></figure><h3>The making: Gemini takes control</h3><p>At the beginning I only had the need. And I want speed, I had not the time to study how to do it, the tool must be done quickly. So Gemini AI gives me:</p><ul><li>The idea to do it with a Browser extension.</li><li>The base of the program and the scaffolding for VSCode.</li><li>How to include an OCR library and use it: Tesseract.</li><li>Performance and security improvements.</li></ul><p>The whole process takes about 6 hours and the most of the time was improving the already made base. In this kind of stuff AI and Gemini is very efective!!</p><p>The result is there <a href="https://github.com/JalmarazMartn/OCREdgeExtension">JalmarazMartn/OCREdgeExtension</a></p><p>You can use it in your browser if you want, downloading the zip, with Code/Download Zip.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/452/1*5ZikzPlfYs-qgOmj0rUBZg.png" /></figure><p>extracting it in your pc and installing in your Browser.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/906/1*lzJZq88tXmvZyVuN3m99WQ.png" /></figure><h3>Basic breakdown and resources</h3><p>The basic of create a Browser extension is doing a JavaScript project with these elements:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/160/1*vG8bWk6J-yo6uU2vaKVo8A.png" /></figure><ul><li>Manifest, for workspace settings.</li><li>Background script to manage the extension.</li><li>An initial script and page. In this workspace: sidebar.html and sidebar.js.</li></ul><p>As extension UI, you can make a popup page or a sidebar page, and additionally a setup page to set the preferences of the extension.</p><p>After that I am improving the extension with the help of this book:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/186/1*QiJwzZgnLJpzus-6AbiPmg.png" /><figcaption><a href="https://www.buildingbrowserextensions.com/">The ultimate Chrome extension tutorial | Building Browser Extensions, 2nd edition</a></figcaption></figure><p>If you read this highly recommended book, you’ll see there’s an entire world behind browser extensions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07232a323b8c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM Tool from get AL Business Central record fields]]></title>
            <link>https://jesusalmarazmartin.medium.com/llm-tool-from-get-al-business-central-record-fields-e800367c4581?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/e800367c4581</guid>
            <category><![CDATA[vscode]]></category>
            <category><![CDATA[d365-business-central]]></category>
            <category><![CDATA[llm-tool]]></category>
            <category><![CDATA[business-central]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 21 Dec 2025 18:34:06 GMT</pubDate>
            <atom:updated>2025-12-21T18:34:06.017Z</atom:updated>
            <content:encoded><![CDATA[<h3>What is an LLM Tool for?</h3><p>As I previously wrote in post <a href="https://medium.com/@jesusalmarazmartin/%EF%B8%8F-how-i-built-a-vscode-llm-tool-to-access-the-clipboard-agent-use-case-4ba0f5fcaa26">🛠️ How I Built a VSCode LLM Tool to Access the Clipboard: Agent use case | by Jesus Almaraz Martin | Medium</a></p><p>a LLM tool for VSCode is a kind of API in VSCode (alike MCP) that allows AI agents doing things they can not do, or can not do very well:</p><ul><li>Access to system capabilities.</li><li>Deterministic operations better solved with a program than an AI process.</li></ul><h3>The plan: add LLM tools to extensions</h3><p>I have the proposal of doing in all my VSCode extensions, LLM tools to allow access AI to some extension features.</p><p>So, in my extension <a href="https://marketplace.visualstudio.com/items?itemName=JessAlmarazMartin.JALVariableNaming">JAL Variable Naming — Visual Studio Marketplace</a> I did inside an LLM tool to bring the list of fields of a record to AI agents. This can ease AI operations, like make automatically assign from a table to another or make AL Datatransfers via agent:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*F_qWyEXFZpUBkblrAj_mPA.png" /></figure><h3>Detailed considerations</h3><p>This tools returns not only regular fields of the record, returns fields from extensions. This is due the technical implementation, and I know, few people will mind this, I take the fields from hover instead symbol definition. Hover includes extension fields, this is the reason for my choice over symbol. And by the way, is faster.</p><p>And yes I know, I could re-invented the wheel because BC AL, will include this kind of things, but I need this now, not many months ago. I hope this could be useful for anyone else.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e800367c4581" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Automate Style Guide Compliance]]></title>
            <link>https://jesusalmarazmartin.medium.com/how-to-automate-style-guide-compliance-cca25b1a7aa9?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/cca25b1a7aa9</guid>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[vscode]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[api]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sun, 30 Nov 2025 19:10:26 GMT</pubDate>
            <atom:updated>2025-11-30T19:10:26.451Z</atom:updated>
            <content:encoded><![CDATA[<h3>Company style guide use case</h3><p>In some company areas we have style guides to ensure text is adapted appropriately. In this era of more and better, is easy to help the following of these style guide: give AI the guide, the text to evaluate, and ask it to evaluate the style compliance.</p><p>But not enough. The good part of this stuff is do thing faster and faster: what about if I integrate this in a management program to set a style guide once, and every time we type a text get an style evaluation in automatic way?</p><p>One use case could be a tech support app with a style guide for task descriptions. Each time a support task description is entered, the system automatically checks it against the guide.</p><p>In the screen we have another example: an app to colect employee suggestion, and before the employee send the suggestion, the system check guide compliance:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/631/1*Ri9PMwU4Tbmh1zlIJ663hg.png" /></figure><p>The checking includes, score 1 to 10 of compliance, improving suggestion, and the biggest style fail of the suggestion:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/843/1*GxE5C1sSZGr5RgoSzVcPVQ.png" /></figure><h3>The internals</h3><p>After typing the text box we call Gemini API with this prompt:</p><pre>        &quot;parts&quot;: [<br>          {<br>            &quot;text&quot;: &quot;Siguiendo esta guía de estilo: &quot; + StyleGuideURL +<br>            &quot;evalúa el siguiente texto y sugiere mejoras. La respuesta debe ser concisa y en JSON, con dos datos: &quot; +<br>            &quot;{CumpleGuia: puntuación cumplimiento de 1 a 10, Sugerencia: Sugerencia de longitud similar a la entrada,MayorError: de forma concisa destaca el mayor error del texto}&quot; +<br>            &quot;Este es el texto a evaluar: &quot;<br>             + textToEvaluate</pre><p>You can find the script in <a href="https://github.com/JalmarazMartn/GeminiAPI/blob/master/StyleGuide.js">GeminiAPI/StyleGuide.js at master · JalmarazMartn/GeminiAPI</a></p><p>It is only a simple API call, without any additional develop kit, this way can be used to be adapted in many programming languages.</p><p>Another point is “StyleGuideURL”. In this variable I pass the URL of a post about corporative language and communication: this is the source of knowledge to the compliance checking.</p><p>Here we have two ways to pass a guide:</p><ul><li>We can pass an URL.</li><li>We pass all the content of the guide in the prompt.</li></ul><p>Looks like best option is pass the URL instead the full text in the prompt, for tokens economy and speed.</p><p>Feature recap. The inputs are:</p><ul><li>Prompt.</li><li>Text to evaluate.</li><li>URL of style guide.</li></ul><p>The outputs:</p><ul><li>Scoring from 1 to 1o of guide compliance.</li><li>Improve suggestion.</li><li>The weakest point of the text to evaluate.</li></ul><p>The quality of the check depend on the style guide and the instructions of the prompts. We can see in the example, that the suggestion and the weakest point is ok, but scoring not so good, is to high for a bad input text.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cca25b1a7aa9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to re-use prompts faster in Visual Studio Code]]></title>
            <link>https://jesusalmarazmartin.medium.com/how-to-re-use-prompts-faster-in-visual-studio-code-99e09503febb?source=rss-4a2b13c05138------2</link>
            <guid isPermaLink="false">https://medium.com/p/99e09503febb</guid>
            <category><![CDATA[github-copilot]]></category>
            <category><![CDATA[vscode]]></category>
            <category><![CDATA[vscode-extension]]></category>
            <dc:creator><![CDATA[Jesus Almaraz Martin]]></dc:creator>
            <pubDate>Sat, 29 Nov 2025 06:19:23 GMT</pubDate>
            <atom:updated>2025-11-29T06:19:23.228Z</atom:updated>
            <content:encoded><![CDATA[<h3>Visual Studio Code settings to re-use prompts</h3><p>The need of speed again: Like many programmers, work often feels like a constant state of urgency — every advantage counts.<br> If you’ve created a great prompt and want to reuse it quickly, here’s how to configure it:</p><ol><li>Press <strong>F1</strong> in VS Code.</li><li>Select <strong>Preferences: Open User Settings (JSON)</strong>:</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/766/1*AkCSpNxMBM9XK-nz9nl4Sg.png" /></figure><p>Then you can edit this setting from Co- Pilot extension:</p><pre>    &quot;chat.promptFilesLocations&quot;: {<br>        &quot;\\Jesus\\Documents\\Proyecto js\\jamcopilotprompthelper\\examples\\&quot;: true<br>    },</pre><p>Saving the prompts in this folder, you can re-use them when you use command “Chat: run prompt”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/350/1*Nkz2R52DKwoeUSNA0_kOAg.png" /></figure><p>The you can choose prompts from the settled folders of “chat.promptFilesLocations”. Important<strong>:</strong> Make sure to save your prompt files with the double extension <strong>.prompt.md</strong>: other way the prompt will not be visible for run prompt command.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/412/1*gUETNS_pMN30o8Cudo-mHQ.png" /></figure><p>Recap: So, with this sequence you can organize the prompt re-use:</p><ul><li>Setting folders in promptFilesLocations.</li><li>Saving new prompts in this folder with double extension <strong>.prompt.md</strong></li></ul><h3>Extension JAM prompt helper</h3><p>In market place I have an extension: <a href="https://marketplace.visualstudio.com/items?itemName=JessAlmarazMartin.jamcopilotpromptmanager">JAM Copilot Prompt Manager — Visual Studio Marketplace</a></p><p>This extension allows you to capture code changes as reusable prompts:</p><ul><li><strong>Start Recording</strong><br> Select the initial code and run: JAM Copilot Prompt Manager: Begin Prompt Recording</li><li><strong>Finish Recording</strong><br> After modifying the code, select the end point and run:JAM Copilot Prompt Manager: End Prompt Recording</li></ul><p><strong>What Happens Next</strong><br> The system will:</p><ul><li>Generate an explanation of the changes.</li><li>Include the code <strong>before</strong> and <strong>after</strong>.</li><li>Save everything in a .prompt.md file inside the folder defined by:jamcopilotpromptmanager.promptsPath</li><li><strong>Important Note</strong><br> If this path is <strong>not included</strong> in chat.promptFilesLocations, you’ll see a warning.<br> Why? Because the goal is to store prompts in one of these folders for easy reuse.</li></ul><p>You can see this process in action in the demo video.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fof1bJzLIBJI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dof1bJzLIBJI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fof1bJzLIBJI%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/5162e788ca31f2e5269b1c28c69634d4/href">https://medium.com/media/5162e788ca31f2e5269b1c28c69634d4/href</a></iframe><p>Hope you found this Helpful.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99e09503febb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>