Looking things up

The most common tool of all is one that simply looks things up. To see why it matters so much, recall two facts we have already nailed down. The model's knowledge is frozen at the moment training stopped, so it knows nothing newer than that. And the only thing it can actually reason over is what sits on the board, the context window.

Put those together and the fix writes itself. If the model cannot know something new, then before it answers, fetch the thing and lay it on the board. Now the model is not straining to recall a half-remembered fact; it is reading a fresh document sitting right in front of it.

That pattern has a clunky name, and placing it in the context window before the model answers, so its reply is grounded in real documents rather than its frozen memory.), and the longer label you may hear is RAG, retrieval-augmented generation. The shape is always the same: a question comes in, a search step finds documents that relate to it, those documents drop onto the board, and the model writes its answer grounded in what it just read instead of what it vaguely remembers.

It does not make hallucination vanish, worth being honest about. The model can still misread the document or stretch what it says. But it tilts the odds hard toward accuracy, especially for anything recent or specialised, because now there is real material on the board to lean on.

And it reveals the deeper truth about that board. It was never just a transcript of the chat. It is a workspace. You can lay anything on it for the model to work over: documents, search results, the output of a tool it just ran, code, a table of numbers. The richer and more relevant what you put on the board, the more capable the agent on top of it. An agent, in the end, is largely the art of getting the right things onto the board at the right time.