Completion: Definition and Examples

In artificial intelligence, a completion refers to the text generated by a language model in response to an input (prompt). The term comes from the fundamental operation of LLMs: they are trained to predict the next token in a sequence, thereby "completing" the text provided to them. Each word or fragment produced results from a probabilistic estimate of what should logically follow.

Historically, early models like GPT-2 and GPT-3 operated exclusively in completion mode: they were given a text beginning and extended it. OpenAI's API even had a dedicated endpoint called "Completions." With the advent of conversational models (ChatGPT, Claude), the paradigm shifted to chat mode, where exchanges are structured as messages (system, user, assistant), but the underlying mechanism remains a form of completion.

The quality of a completion directly depends on the quality of the prompt. A vague prompt will produce a generic completion, while a precise, structured, and contextualized prompt will generate a relevant and targeted response. This direct relationship underpins the entire discipline of prompt engineering.

Several parameters influence completion: temperature controls the randomness or determinism of the response, max_tokens limits its length, and frequency or presence penalties modulate the diversity of the vocabulary used. Understanding these levers allows obtaining completions suited to each use case, whether for creative writing, data analysis, or code generation.