A text model generates a reply from learned parameters and the context supplied for that request. That context can include instructions, messages, selected conversation history, files and tool results. Your newest message is one part of the input.
Tokens, context and learned patterns
Text is represented as tokens, which can be words, parts of words or punctuation. The model generates successive output tokens using patterns learned during training. It does not retrieve a single stored answer for every question. OpenAI’s token explanation describes how these units differ from words.
The context window limits what a model can process in a request. A product may select or summarize earlier messages rather than include an entire conversation. A supplied document adds information to the request; it does not by itself retrain the model. Treat a missing source passage as missing input, even when the assistant produces a confident answer.
Example: identify what supports the answer
Use these example notes: “The library workshop starts at 10:00 on Saturday. Bring a laptop. Booking is required. The room has not been assigned.” Our task is to make a two-item preparation checklist. The source contains an unknown room and no booking link.
Make a two-item preparation checklist using only Source A.
Keep unknown details unknown. Add one clarification question.
Source A: The library workshop starts at 10:00 on Saturday.
Bring a laptop. Booking is required. The room has not been assigned.Example answer
- Book a place for the Saturday workshop starting at 10:00.
- Bring a laptop.
Clarification: Which room will host the workshop?
The instructions determine the format; Source A supports the time, booking requirement and laptop. A room number, instructor’s name or booking URL would need another source. None can be justified by the model’s fluent wording.
Retrieval and memory are separate inputs
A search-enabled assistant can add retrieved material. OpenAI’s web-search documentation describes that tool-based access. Asking for “today’s room” does not establish that a search happened: inspect the tool result and open the relevant source.
Product memory can also supply preferences or details across chats. ChatGPT’s memory controls govern that feature; it is distinct from the model’s training and from the text you paste into one request.
Check the input before adding more
Ask which information the task needs, whether it is present and whether it is current. Include the relevant passage and remove conflicting drafts or unrelated background. In the context lesson, you will practice choosing evidence rather than simply making a prompt longer.