Skip to main content
Prompt engineering is the art and science of crafting effective instructions for AI models, particularly Large Language Models (LLMs). A well-designed prompt can significantly enhance the quality, relevance, and safety of AI-generated responses. This guide will walk you through the key concepts and best practices in prompt engineering.

​

Rememeber to always include the variables {kb_context} and {about_context} in your prompt or else the agent wont know the retrieved chunks from the RAG.

​
Key Concepts

​

Chain of Thought (CoT) reasoning is a technique that involves breaking down complex problems into a series of intermediate steps. This approach helps the AI model to:
  1. Understand the problem more thoroughly
  2. Show its reasoning process
  3. Arrive at more accurate conclusions
Example:

Human: What's the result of 25 * 18?   


AI: Let's approach this step-by-step:   

 
1. First, let's break down 25 * 18 into (20 + 5) * 18   


2. Now we can calculate:   


   a) 20 * 18 = 360   


   b) 5 * 18 = 90   


3. Adding these results:   


   360 + 90 = 450   


Therefore, 25 * 18 = 450
Few-shot learning is a technique where you provide the AI with a small number of examples to guide its understanding of the task. This can be particularly useful when you want the AI to follow a specific format or style in its responses.
  • One-shot learning: Providing one example
  • Two-shot learning: Providing two examples
  • Few-shot learning: Providing a few (typically 3-5) examples
Example:


Human: Translate the following English phrases to French. Here are two examples:   

 


English: Hello, how are you?   

French: Bonjour, comment allez-vous ?   

 

English: Where is the nearest restaurant?   

French: Où est le restaurant le plus proche ?   



Now, translate this:   

 
English: I would like to book a hotel room.   




AI: Based on the examples provided, here's the translation:   



 
English: I would like to book a hotel room.   

French: Je voudrais réserver une chambre d'hôtel.
⌘I