- „Who„/What do you expect the LLM to be like?
- Otherwise, the LLM will prioritise the most probable data for general knowledge.
- ’You are an expert researcher.’
- What is the task?
- Try to be as specific as possible to avoid hallucinations.
- ’Find the number of family-owned wineries in France in 2014.’
- What is the context of the task?
- LLMs are like unformed, eager to help young employees in your team. If you don’t create obvious boundaries, it can derail the solution completely. Also, sometimes, with the right context, an LLM can suggest something that wouldn’t have occurred to you.
- ’Based on this data, we will decide whether to invest in new vineyards in Poland to compete with these French vineyards, so it’s important to have a confidence in the data for this decision.„
- What is the expected result of your query, and what would be an undesirable outcome?
- LLMs can hallucinate at every stage, and they love to provide a too long answers.
- „The ideal result would be a bullet-point list with sources and links, showing the total number of French wineyards with a percentage indicating the completeness of the list. A bad result would look like, 'The number of vineyards is between 10 and 1,000, based on multiple sources which I don’t list and haven’t double-checked’.’”
- What would be the desired sources for the query and what would be unacceptable sources?
- LLMs love to create sources of their own or combine multiple ones and present something that doesn’t exist.
- ’Ideal sources would be high-quality data from the French government and French vineyard organisations, and unacceptable sources would be forum or Reddit posts without any sources other than anonymous accounts.’
- What is the emotional context of this query?
- It’s crazy, but it helps to prioritise the work for the LLM.
- ’This is extremely important work, so make sure you complete the task with extreme focus and make no mistakes; otherwise, I will be very angry with you.’
- Experiment with the prompt at least once.
- The difference between two similar prompts can be mind-blowing, as LLMs are probabilistic, not deterministic. Try it at least once before settling on a solution.
- Don’t hesitate to redo the answer if you’re not happy with the LLM’s first attempt.
Dodaj komentarz