[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"post-how-to-orchestrate-language-models-in-practice":3,"$fXeC77ibo_YuG-ZnFG3hwyaNC83zFAKw-FvBNoL2JrIM":15},{"slug":4,"title":5,"excerpt":6,"category":7,"author":8,"readingTime":9,"coverImage":10,"bodyHtml":11,"metaTitle":12,"metaDescription":13,"date":14},"how-to-orchestrate-language-models-in-practice","Orchestrating language models: from one prompt to a system","One prompt is a demo. An application is something else. On splitting work, routing between models, and where to leave ordinary code alone.","Guides","MightCore","9 min",null,"\u003Cp>The first AI feature is usually one big prompt. You put everything in it, the model somehow copes, and it feels easy.\u003C\u002Fp>\n\u003Cp>The second version of that prompt is three hundred lines, contains eight instances of \"IMPORTANT:\", and still returns nonsense now and then. That is usually when someone asks whether a better model would help. Normally it would not. The problem is not the model. The problem is one task doing five things at once.\u003C\u002Fp>\n\n\u003Ch2>Step one: break the task apart\u003C\u002Fh2>\n\u003Cp>Take processing an incoming invoice. As one prompt it reads: \"read the invoice, verify the supplier, check the amounts, assign the right category and write a summary\".\u003C\u002Fp>\n\u003Cp>Split into steps it looks different:\u003C\u002Fp>\n\u003Col>\n\u003Cli>extract the fields from the document (supplier, registration number, total, date, line items),\u003C\u002Fli>\n\u003Cli>look the supplier up in our database by registration number — \u003Cem>a plain SQL query, no model\u003C\u002Fem>,\u003C\u002Fli>\n\u003Cli>check the line items add up to the total — \u003Cem>arithmetic, no model\u003C\u002Fem>,\u003C\u002Fli>\n\u003Cli>suggest a category based on this supplier's history,\u003C\u002Fli>\n\u003Cli>if anything is uncertain, hand it to a person.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>Suddenly that is two model calls instead of one, but each has a single clear job. Steps two and three need no AI at all — and that is the more important half. A model \"verifying\" sums is the most expensive way ever devised to get addition wrong.\u003C\u002Fp>\n\u003Cp>If you take one sentence from this article: \u003Cstrong>use AI where you cannot write the rules. Everywhere else, write the rules.\u003C\u002Fstrong>\u003C\u002Fp>\n\n\u003Ch2>Chaining, and what it costs\u003C\u002Fh2>\n\u003Cp>Put the steps in sequence and one step's output feeds the next. It works, with two catches.\u003C\u002Fp>\n\u003Cp>First: errors do not add up, they multiply. Nine steps at ninety percent reliability does not give you ninety percent. It gives you about thirty-eight. So keep the chain short and check after each step that the output makes sense at all.\u003C\u002Fp>\n\u003Cp>Second: latency. Every step is a network call measured in seconds. Five in a row means a user is waiting. Run independent steps in parallel, and anything that can happen in the background should happen in the background.\u003C\u002Fp>\n\n\u003Ch2>Routing: not every task needs the strongest model\u003C\u002Fh2>\n\u003Cp>Models differ in price and speed, often by an order of magnitude. Sending \"is this a complaint or an enquiry?\" to your most expensive model is taking a lorry to the corner shop.\u003C\u002Fp>\n\u003Cp>A sensible split looks roughly like this: simple, high-volume work (classification, extraction, short summaries) goes to a small fast model. Work that needs reasoning, joining several things together, or writing code goes to a strong one. Plus one rule for safety: when the small model is unsure, let it escalate.\u003C\u002Fp>\n\u003Cp>In practice this means having one place in the code that decides which model gets a task. Not ten places with a model name hard-coded next to a prompt. The second one catches up with you the first time pricing changes, or the first time a new model ships.\u003C\u002Fp>\n\n\u003Ch2>Have the model return data, not prose\u003C\u002Fh2>\n\u003Cp>This is the cheapest improvement available. Ask for structured output — JSON against a schema — instead of a sentence you then fish a number out of with a regex.\u003C\u002Fp>\n\u003Cp>The reason is not tidiness. Structured output can be \u003Cem>validated\u003C\u002Fem>. Field missing? Rejected. Amount not a number? Rejected. Category not in the allowed list? Rejected. You now have a gate between the model and your database that stops a hallucination before it breaks something.\u003C\u002Fp>\n\u003Cp>And when output fails validation, try again. Models are non-deterministic — the second attempt is often fine.\u003C\u002Fp>\n\n\u003Ch2>Where it actually falls over\u003C\u002Fh2>\n\u003Cp>A few things that always show up in production and never in a demo:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>External APIs go down.\u003C\u002Fstrong> Set timeouts and have a fallback. An AI feature that takes checkout with it because a provider had an outage is worse than no AI feature.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Costs grow quietly.\u003C\u002Fstrong> Three cents a call is nothing until it is a hundred thousand calls a month. Measure from day one.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A prompt is input, not code.\u003C\u002Fstrong> Once you interpolate text from a user or an email into a prompt, assume somebody will try to talk you into ignoring your previous instructions. Keep instructions separate from data, and never give a model authority you would not give a stranger.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Cache.\u003C\u002Fstrong> The same question repeats more often than you would think.\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Ch2>When not to do any of this\u003C\u002Fh2>\n\u003Cp>If a condition in code solves the task, solve it with a condition in code. It is cheaper, faster, testable, and it does not get things wrong.\u003C\u002Fp>\n\u003Cp>That sounds obvious, and yet we have seen enough projects where a language model was deciding things an \u003Ccode>if\u003C\u002Fcode> could have handled. Orchestration is not about how many models you wire in. It is about how many you left out because they were not needed.\u003C\u002Fp>","How to orchestrate LLMs — from prompt to working system","A practical look at orchestrating language models: chaining steps, routing between models, validating output, and where not to use AI at all.","2026-03-04T00:00:00.000Z",[16,22,28,33,39,46,53,58,62,68,74,79,84,89,94,99,100,105,111,117,122,127,132,137,142,147,152,157,162,167,172,177,182,187,191,196,201],{"slug":17,"title":18,"excerpt":19,"category":7,"author":8,"readingTime":20,"coverImage":10,"date":21},"how-to-talk-to-a-language-model-prompting-in-practice","How to talk to a model: prompting without incantations","Prompt engineering is not a list of magic phrases. It is the ability to say exactly what you want — which is harder than it sounds.","7 min","2026-07-08T00:00:00.000Z",{"slug":23,"title":24,"excerpt":25,"category":26,"author":8,"readingTime":20,"coverImage":10,"date":27},"how-to-choose-a-language-model-and-control-costs","Choosing a language model and keeping costs under control","Benchmarks tell you almost nothing useful. What actually decides a model choice, and where the costs nobody predicted come from.","Engineering","2026-06-30T00:00:00.000Z",{"slug":29,"title":30,"excerpt":31,"category":26,"author":8,"readingTime":9,"coverImage":10,"date":32},"programming-in-the-age-of-ai-what-actually-changed","Programming in the age of AI: what actually changed","No, it did not replace developers. But it changed where they spend their time — and not all of those changes are comfortable.","2026-06-17T00:00:00.000Z",{"slug":34,"title":35,"excerpt":36,"category":37,"author":8,"readingTime":20,"coverImage":10,"date":38},"ai-modelka-pre-vas-brand-sprievodca","An AI model for your brand: the complete guide","From brief through avatar creation to the first campaign — step by step.","AI UGC","2026-06-16T00:00:00.000Z",{"slug":40,"title":41,"excerpt":42,"category":43,"author":8,"readingTime":44,"coverImage":10,"date":45},"pripadova-studia-cleago","Case study: Cleago — a platform built on context","How we designed and built a solution for Cleago (www.cleago.sk) by first understanding the context and only then coding.","Case studies","5 min","2026-06-02T00:00:00.000Z",{"slug":47,"title":48,"excerpt":49,"category":50,"author":8,"readingTime":51,"coverImage":10,"date":52},"where-not-to-use-ai-and-why-nobody-tells-you","Where not to use AI (and why nobody tells you)","A company selling AI has little incentive to talk about its limits. Let us fix that — these are the places AI simply does not belong.","For business","6 min","2026-05-27T00:00:00.000Z",{"slug":54,"title":55,"excerpt":56,"category":43,"author":8,"readingTime":44,"coverImage":10,"date":57},"pripadova-studia-produktove-fotky","Case study: 80% less time spent creating product photos","A real example of deploying AI photos in an online store — from brief to results.","2026-05-19T00:00:00.000Z",{"slug":59,"title":60,"excerpt":61,"category":37,"author":8,"readingTime":44,"coverImage":10,"date":57},"ai-model-consistent-brand-across-campaigns","A consistent brand with an AI model across campaigns","An AI model can be a brand's steady face — if you handle consistency and transparency the right way. Here is how.",{"slug":63,"title":64,"excerpt":65,"category":50,"author":8,"readingTime":66,"coverImage":10,"date":67},"where-ai-actually-fits-in-a-company","Where AI actually fits in a company (and where it is just flashy)","Concrete applications across departments — what works today, what needs preparation, and what is still a demo rather than a tool.","8 min","2026-05-06T00:00:00.000Z",{"slug":69,"title":70,"excerpt":71,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":73},"rest-vs-graphql-pre-eshopy","REST vs. GraphQL API for modern online stores","When to choose which approach and what the impact on performance and development is.","Development","2026-04-21T00:00:00.000Z",{"slug":75,"title":76,"excerpt":77,"category":50,"author":8,"readingTime":9,"coverImage":10,"date":78},"how-to-bring-ai-into-your-company-first-steps","How to bring AI into your company without wasting the money","Most AI projects do not fail on technology. They fail because nobody said what was supposed to get better. Where to start instead.","2026-04-15T00:00:00.000Z",{"slug":80,"title":81,"excerpt":82,"category":43,"author":8,"readingTime":51,"coverImage":10,"date":83},"case-study-monolith-to-modular-migration","Case study: from a monolith to a modular architecture with no downtime","An illustrative example of gradually modernising an older application — where every change was risky and maintenance expensive.","2026-04-14T00:00:00.000Z",{"slug":85,"title":86,"excerpt":87,"category":7,"author":8,"readingTime":66,"coverImage":10,"date":88},"ai-agents-and-tool-calling-when-it-makes-sense","AI agents and tool calling: when it makes sense and when it does not","An agent is a model allowed to act. That is interesting and dangerous in equal measure. How it works, and where to be careful.","2026-03-25T00:00:00.000Z",{"slug":90,"title":91,"excerpt":92,"category":7,"author":8,"readingTime":51,"coverImage":10,"date":93},"context-driven-development-context-gathering-in-practice","Context Driven Development in practice: how gathering context changes the outcome","The most expensive mistakes come from a misunderstood brief. Here is what the context gathering that prevents them looks like.","2026-03-17T00:00:00.000Z",{"slug":95,"title":96,"excerpt":97,"category":98,"author":8,"readingTime":51,"coverImage":10,"date":93},"uctovnictvo-novej-generacie","Next-generation accounting: a platform built on context","A vision of an intelligent layer on top of existing accounting tools.","Accounting",{"slug":4,"title":5,"excerpt":6,"category":7,"author":8,"readingTime":9,"coverImage":10,"date":14},{"slug":101,"title":102,"excerpt":103,"category":98,"author":8,"readingTime":44,"coverImage":10,"date":104},"digitalising-accounting-e-invoicing","Digitalising accounting: e-invoicing and what it brings","Electronic invoicing and reporting are becoming the standard. What it means for businesses and how to prepare without panic.","2026-02-17T00:00:00.000Z",{"slug":106,"title":107,"excerpt":108,"category":109,"author":8,"readingTime":66,"coverImage":10,"date":110},"what-is-an-llm-large-language-model-explained","What an LLM is: large language models without the mystique","How does a program that predicts the next word end up writing working code? A look at what actually happens inside a language model.","Fundamentals","2026-02-11T00:00:00.000Z",{"slug":112,"title":113,"excerpt":114,"category":115,"author":8,"readingTime":44,"coverImage":10,"date":116},"ako-ai-setri-naklady-na-video","How AI cuts the cost of video content production","Concrete numbers and a workflow for creating AI videos for online stores.","Marketing","2026-02-10T00:00:00.000Z",{"slug":118,"title":119,"excerpt":120,"category":109,"author":8,"readingTime":20,"coverImage":10,"date":121},"what-is-artificial-intelligence-explained-without-the-marketing","What artificial intelligence is (and what it is not)","The word AI now means everything, and therefore nothing. Here is what is actually behind it, and where technology ends and marketing begins.","2026-01-21T00:00:00.000Z",{"slug":123,"title":124,"excerpt":125,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":126},"vector-databases-and-embeddings","Vector databases and embeddings: how machines grasp meaning","Semantic search sits behind many AI features. Here is what embeddings are and why modern data work rests on them.","2026-01-20T00:00:00.000Z",{"slug":128,"title":129,"excerpt":130,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":131},"trendy-v-ai-vyvoji-2026","Trends in AI development for 2026","What awaits companies in the area of AI agents, automation and infrastructure.","2026-01-14T00:00:00.000Z",{"slug":133,"title":134,"excerpt":135,"category":43,"author":8,"readingTime":44,"coverImage":10,"date":136},"case-study-ai-product-photography-cosmetics","Case study: AI product photography for a cosmetics e-shop","An illustrative example of how AI content replaced repeated photoshoots and brought a consistent visual identity across seasons.","2025-12-09T00:00:00.000Z",{"slug":138,"title":139,"excerpt":140,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":141},"integrating-ai-into-existing-systems","How to integrate AI into existing systems without a rewrite","You do not have to throw away working software to use AI. Here is an approach that adds value step by step, with low risk.","2025-11-18T00:00:00.000Z",{"slug":143,"title":144,"excerpt":145,"category":7,"author":8,"readingTime":20,"coverImage":10,"date":146},"gdpr-a-ai-obsah","GDPR and AI content: what to watch out for","The legal minimum for companies working with AI content and personal data.","2025-11-11T00:00:00.000Z",{"slug":148,"title":149,"excerpt":150,"category":37,"author":8,"readingTime":51,"coverImage":10,"date":151},"virtualne-ai-modelky","Virtual AI models: the future of advertising or a passing trend?","The possibilities, limits and ethics of virtual influencers for brands.","2025-10-20T00:00:00.000Z",{"slug":153,"title":154,"excerpt":155,"category":37,"author":8,"readingTime":44,"coverImage":10,"date":156},"ai-ugc-in-performance-marketing","AI UGC in performance marketing: what works and what applies","How to use AI content in Meta and TikTok campaigns, why creative testing matters, and what AI-labelling rules apply.","2025-10-14T00:00:00.000Z",{"slug":158,"title":159,"excerpt":160,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":161},"llm-hallucinations-how-to-limit-them","LLM hallucinations and how to limit them in practice","Why AI sometimes states nonsense with confidence, and the techniques we use to keep output trustworthy.","2025-09-16T00:00:00.000Z",{"slug":163,"title":164,"excerpt":165,"category":37,"author":8,"readingTime":44,"coverImage":10,"date":166},"co-je-ai-ugc","What AI UGC is and why the whole world is talking about it","An introduction to AI-generated UGC and its impact on advertising and customer trust.","2025-09-15T00:00:00.000Z",{"slug":168,"title":169,"excerpt":170,"category":43,"author":8,"readingTime":51,"coverImage":10,"date":171},"case-study-b2b-eshop-faster-delivery","Case study: a B2B e-shop ready in weeks, not months","An illustrative example of how context gathering and AI execution shortened a wholesale e-shop build — without cutting quality.","2025-08-19T00:00:00.000Z",{"slug":173,"title":174,"excerpt":175,"category":115,"author":8,"readingTime":44,"coverImage":10,"date":176},"ai-v-marketingu-od-experimentu-k-vysledkom","AI in marketing: from experiment to real results","How to move from \"playing with AI\" to a measurable return on investment.","2025-08-06T00:00:00.000Z",{"slug":178,"title":179,"excerpt":180,"category":98,"author":8,"readingTime":44,"coverImage":10,"date":181},"ai-invoice-processing-in-accounting","AI invoice processing: from scan to posting","Intelligent document processing cuts the routine of retyping invoices. How it works and where AI has its limits.","2025-07-15T00:00:00.000Z",{"slug":183,"title":184,"excerpt":185,"category":7,"author":8,"readingTime":51,"coverImage":10,"date":186},"rag-why-context-decides-ai-quality","RAG: why context decides the quality of AI output","Retrieval-Augmented Generation connects a language model to your own data. Here is how it works and when to use it.","2025-06-18T00:00:00.000Z",{"slug":188,"title":189,"excerpt":190,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":186},"shopsys-vs-vlastne-riesenie","ShopSys vs. a custom build: when a framework pays off","A decision framework for e-shop owners facing the choice of platform.",{"slug":192,"title":193,"excerpt":194,"category":115,"author":8,"readingTime":44,"coverImage":10,"date":195},"ako-ai-meni-ecommerce-na-slovensku","How AI is changing e-commerce","Practical examples of AI in product content, search and personalisation for online stores.","2025-05-21T00:00:00.000Z",{"slug":197,"title":198,"excerpt":199,"category":72,"author":8,"readingTime":20,"coverImage":10,"date":200},"context-driven-development-novy-pristup","Context Driven Development: a new approach to building software","An explanation of the CDD methodology from gathering context to deployment — step by step.","2025-04-09T00:00:00.000Z",{"slug":202,"title":203,"excerpt":204,"category":72,"author":8,"readingTime":51,"coverImage":10,"date":205},"koniec-ery-predrazeneho-vyvoja","Why the era of overpriced software development is over","How AI and a context-driven approach are changing the economics of building software — and why paying for inflated hours no longer makes sense.","2025-03-12T00:00:00.000Z"]