LLM

Understanding one mechanism removes most of the misconceptions. The model does not know the answer and does not look it up: it predicts the next fragment of text in sequence, based on statistical patterns in its training data and on what has already been written in the context. Hence hallucinations, because a plausible continuation and a true continuation are different things and the model is optimised for the first. Hence the context window limit, since everything the model remembers in a conversation is the text that physically fits into the input. And hence the sensitivity to phrasing, because different framings land in different regions of the distribution.

For a Russian project the choice of model is constrained by law as well as quality. When personal data of Russian users enters the model, meaning names, phone numbers or the content of enquiries, the requirements of the personal data law apply, including database localisation. That leaves three workable scenarios: domestic services such as GigaChat and YandexGPT, open-weight models deployed inside a Russian perimeter, and foreign APIs restricted to data that is not personal. Settling this in the first week of a project costs far less than settling it after launch, when moving means rewriting the data layer.

The practical conclusion for a business: LLMs handle tasks where variation is acceptable and a reviewer exists, and handle badly anything that needs a deterministic exact answer. Draft copy, ticket classification, field extraction from documents, conversation summarisation and first-line support against a knowledge base are working scenarios. Order price calculation, legally binding wording and answers about stock levels belong to ordinary code and databases, with the model at most phrasing the result. The most common deployment mistake is handing the model work that had to be deterministic.

Related terms

• RAGRAG is a pattern where a language model first retrieves relevant fragments from your own knowledge base and answers from them, which delivers current data and a source citation without any fine-tuning of the model itself.
• EmbeddingsEmbeddings are numeric vectors a model produces from text, images or products so that items close in meaning end up close in the vector space, which is what makes semantic search, recommendations and duplicate detection possible.
• Prompt engineeringPrompt engineering is the design of instructions for a language model: role, context, task, output format and constraints are written so the result is reproducible, and quality is judged against a set of test cases rather than by impression.
• AI agentAn AI agent is a program in which a language model does not merely answer but acts: it receives a goal, chooses its own tools such as search, APIs, a database or sending mail, works through steps in a loop and stops when the task is done or a limit is hit.
• Russian personal data law (152-FZ)152-FZ is the Russian personal data law: it requires that data on Russian citizens be collected in databases located in Russia, that the regulator be notified of processing, that consent be obtained separately, and that any breach be reported within 24 hours.

Related services

• AI and ML developmentWe build AI that solves a defined task and pays for itself, not demos for the sake of demos. Classifiers, recommendations, text and document processing, LLM assistants, and model integration into an existing product.
• IT consulting and product auditThe most expensive mistakes in software happen before the first line of code: a misread problem, a stack chosen out of a contractor’s habit, and a specification that does not exist. Consulting exists to settle all of that before the development meter starts running: what to build, out of what, at what cost and in which order. The result is a document, not an opinion on a call.

Read more

• An AI Assistant on Your Own Knowledge Base: How RAG Works and What It CostsHow a RAG assistant differs from a button-tree bot and from an LLM with no data, where hallucinations come from and what actually reduces them, what SaaS costs against a custom build, and which month the project breaks even.
• Where AI and ML Actually Pay Off in a Business, and Where They Are Expensive HypeMachine learning pays off where there is a repeated decision, a lot of similar data and tolerance for error. Missing any one of the three means do not. Here are the use cases that work, their realistic accuracy, the running costs and how to run a four to six week pilot.
• GEO in 2026: how to get cited by ChatGPT, Alice and AI OverviewsGEO is not a separate channel, it is a layer on top of SEO. Here is how six answer engines pick sources, what makes a page quotable, the honest truth about llms.txt, and a weekly method for measuring AI visibility.