Context Machine

A language program like ChatGPT only knows what is in its request, and what it absorbed while learning from texts. Everything else is unknown to it: not your school notes, not yesterday’s invoice, not this morning’s message. A Context Machine is the layer in front of it that procures exactly this missing material. It searches databases, files, and the web, selects the relevant passages, and pushes them to the model together with the question. The name emphasizes that here it is not the model doing the main work, but the machinery around it. In the industry, people therefore now speak of context engineering rather than just cleverly phrased questions.

Why good models fail on bad material

Two companies can buy the same language model. Yet one delivers usable answers and the other nonsense. The difference is almost never the model, but the material it receives. A model without access to the right documents guesses, and unfortunately guessing looks very convincing with language models.

On top of that comes a hard technical limit. Every model has a context window, i.e. an upper bound on how much text it can process at once. Even large windows hold only a fraction of a company archive. So one has to select. And the selection determines the quality of the answer.

A second reason is money. Providers charge based on the amount of text going in and out. Anyone who sends along a hundred pages every time pays for a hundred pages, even if only one paragraph was relevant. A good Context Machine therefore saves not just errors, but also costs. More context is thus not automatically better.

From searching to editing together

The first step is breaking things down. Long documents are cut into sections of a few hundred words. Each section receives a series of numbers describing its meaning, a so-called embedding. Texts with similar content receive similar number series, even if they use different words.

When a question comes in, it is translated into the same numerical form. The system then searches out the sections with the most similar number series. This procedure is called Retrieval Augmented Generation, or RAG for short. Many systems combine it with classic keyword search, because pure meaning-based search often misses the mark with product numbers or names.

After that comes the part that distinguishes a Context Machine from a mere search. The hits are re-sorted, duplicates are removed, outdated versions are filtered out. Added to this are further building blocks: the conversation history so far, fixed rules for the model’s behavior, and results from tools such as calculators or calendar queries. In long conversations, the older part is summarized so that it takes up less space. Only this finished, assembled package then goes to the model.

Where this layer is already at work

Every chatbot on a company website that knows its own prices and delivery times runs on such a layer. The same applies to coding assistants that know your project’s code, even though that code was never part of their training. The feature of uploading a PDF and asking questions about it is, at its core, also a small Context Machine.

In business news, the term mostly appears in connection with company knowledge. Banks, law firms, and government agencies want to use AI without handing their data over into someone else’s training. A Context Machine is the usual way to do this: the data stays in the company’s own database, the model only sees it briefly for a single answer.

A common misconception is that huge context windows make this technique obsolete. In practice, the opposite is observed. The more text a model receives at once, the more likely it is to overlook important things in the middle. Selection thus remains necessary — only the boundary shifts.