How does an LLM process a user’s prompt?

How does an LLM process a user’s prompt?

An LLM processes a prompt through several stages that transform text into numerical inputs and then transform model outputs back into text. The visible question may also be combined with earlier messages or application instructions before the model receives it.

Preparing the input sequence

Suppose the user asks, “Summarize this paragraph in one sentence.” The application formats the conversation according to the model’s expected template. A tokenizer converts the resulting text and special markers into token IDs. The exact input therefore depends on both the user’s message and the surrounding application.

Each ID selects a learned embedding. Position information helps the network distinguish the order of the tokens. These steps create the numerical representation on which the model operates; the tokenizer itself does not write the summary.

Combining the available context

Transformer layers repeatedly update token representations using attention and other learned transformations. In a standard causal language model, a position can use preceding positions and itself, while later positions are masked. This allows the network to build a representation suitable for predicting a continuation without accessing future answers.

For the summary request, the model can use the paragraph, the requested length, and any relevant earlier instructions that fit within its available context.

Generating the continuation

The output layer produces scores for vocabulary items. A decoding strategy chooses the next token, adds it to the sequence, and repeats until a stopping condition is reached. The application decodes those IDs for display. Retrieval, calculators, or other tools may add extra steps, but they are separate capabilities surrounding the core language-generation process.