Prompt structure

Role ​

• As a [whatever, tbh, Joe Rogan, React Developer, JSON generator, you name it]
• settles the conversation in the context

Instruction ​

• sets the model's way of doing, thinking.
• provides rules
• should contain the information about date, location, summary of previous conversations (if needed for handling the instruction properly)
• prone to knowledge-cutoff

Context ​

• includes a set of data that can be manually provided, generated by the model, or dynamically added by the application's logic
• should be clearly separated from the rest of the prompt to avoid misinterpretation
  • e.g. using ### or any other non-expected separator
  • without separation parts of the context could be mis-understood as the parts of the instruction
• can be generated by the model as a result of reflexion or a chain of thought
• unnecessary noise should be avoided
• length of the context should be optimal to avoid generating excessive costs
• context should be described, e.g., as `next meetings in the calendar [...]``.
• if the context does not contain sufficient information, this should be indicated that:
  • the model should use its own knowledge base
  • the model should redirect to human contact
  • etc.
• prompts can be composed (one prompt for summarising the website, another one to deal with the information)
  • in such cases the combined prompts should be well-defined to avoid the unwanted behavior (due to undeterministic nature of LLMs)
    • if one prompt fails, the subsequent ones will fail too (please, use try..catch)

Examples ​

• guide the behavior of given LLM
  • focus
  • speech style
  • response format
  • dataset classification
  • etc.
• LLMs can learn from patterns and subtle changes in the examples provided (few-shot-learning, in-context-learning)
• Examples can be instructions at the same time
  • Respond in JSON format: {"name": "Krystian", "email": "krystian@example.com"}. [...]

Question ​

• query that the model is to process, eg.
  • a data set for transformation,
  • a question,
  • a command,
  • a simple message as part of a longer conversation
• might be considered by the model as an instruction to execute, which may not be expected in some cases

Completion (answer) ​

• not a part of the prompt itself, but considered as the part of the token-window
• if we are working with the Chat, the answer will be passed for the subsequent chat messages (the whole conversation is)