Basic Principles
- Write clear and specific instructions (AS CLEAR AS POSSIBLE)
- Use delimiters to clearly indicate different parts of input (quotations, tags, etc) - this also helps with
prompt injection - Ask For a structured Output (like HTML/JSON possibly)
- Check whether conditions are satisfied
- Few Shot Prompting
- Use delimiters to clearly indicate different parts of input (quotations, tags, etc) - this also helps with
- Give model time to think
- Make the model work out the answer first, specify the steps required to complete the tasks
- Instruct the model to work out on its own solution before rushing to conclusion
Iterative Prompt Development
- Idea
- Implementation
- Experimental result -> Idea again
Dealing with Hallucinations
- First find relevant information then answe the question based on relevant information using tools like langchain / crewAI
Settings
- Temparature - (0 - 2) - determines confidence of the model in it’s top prediction. 0 means full confidence, ensuring predictability of the model.
- Top P
- Stop Sequence
- Frequency Penalty
- Presence Penalty