Mastering the Art of Prompting: A Creative Guide to Generative AI (Part 3 of 5)

Introduction

The manner in which we interact with AI systems, such as ChatGPT, Claude, or Gemini, significantly influences our user experience. There are fundamental differences between conventional search mechanisms and generative AI models. These differences determine the extent to which one can optimally use these systems. The role of effective prompting is crucial in this context.

Prompt Engineering is a captivating field that intersects artificial intelligence (AI) and natural language processing (NLP). It involves formulating effective prompts or instructions that guide AI models, particularly Large Language Models (LLMs), to produce desired outcomes.

GPT vs Traditional Search Engines

In the evolving landscape of information retrieval, two distinct approaches have emerged: Generative Pre-trained Transformers (GPTs) and Traditional Search Engines. This comparison chart deep dives into the specific nuances of each method, evaluating their performance across various criteria such as retrieval approach, user experience, and fact-checking capabilities. By contrasting the conversational nature of GPTs with the keyword-based mechanics of traditional search engines, we gain an understanding into their respective strengths and limitations. The chart serves as a guide for understanding how these technologies shape our search experiences and the accuracy of the information we receive.

Basic Prompt Structure

This is the basic structure of any prompt we may give and keeping or removing components from the below structure governs the kind of prompting technique it becomes.

• Goal: This is the main task we want to achieve through AI.
• Context: This is the additional data or background details that can guide the model to produce more accurate responses.
• Expectation: This is the desired length, tone or style of the output
• Source: Any additional data that we can provide as a sample for model to learn off of.

Let's explore some basic concepts before understanding the various prompting techniques.

What are Tokens in Prompting?

Tokens are crucial in prompting as they transform complex, variable-length text into a uniform format that the model can understand and process. Each token is analyzed in context, so the model learns not just the token, but its relationship with others. For example, in the sentence “The cat sat on the mat”, each word is a token, and the model learns the meaning of each word in the context of the sentence.

What is the "temperature" setting in Prompting?

The concept of a “temperature” setting is commonly used in AI models, especially those involved in generating sequences, such as text. However, it’s not a universal feature across all AI systems.

In the context of AI, the “temperature” is a parameter that influences the randomness of the AI’s output. It’s like a dial that adjusts the balance between predictability and creativity.

Prompting Techniques

Prompting techniques are language strategies to guide language models towards desired outputs. Understanding these techniques is crucial as they shape how we interact with AI, making it more intuitive and effective. To draw an analogy, think of prompting as a traveler in a foreign land with limited knowledge of the local language. They use simple words, gestures, or even drawings to communicate their needs or ask for directions. Similarly, in prompt engineering, we use carefully crafted prompts to ‘communicate’ with the AI, guiding it to produce the desired output.

Zero-Shot Prompting

Zero-shot prompting involves providing a prompt that is not part of the example data to the model, but the model can generate a result that you desire from its own training.

For example, if you ask the AI “Write a short story about a brave night”, even though this exact question was not part of the training data, the AI can provide an answer like “Once upon a time, in a kingdom far, far away, there lived a brave knight named Sir Gallant. Sir Gallant was known throughout the land for his courage and valor…”. This is an example of zero-shot prompting. The model has not seen any examples of how to write a story about a brave knight from you, but it’s able to generate one based on its pre-existing knowledge and training data. The model uses its understanding of language and the world, learned during pre-training, to generate a response.

Few-Shot Prompting

Few-shot prompting is a technique that enables in-context learning where we provide demonstrations in the prompt to steer the model to better performance. For example, you want it to generate a joke. Instead of just asking it to “tell a joke”, which is zero-shot, you provide it with a few examples of the kind of jokes you want. This is known as few-shot prompting. Below is a sample:

Example 1: “Why don’t scientists trust atoms?” “Because they make up everything!”

Example 2: “Why was the math book sad?” “Because it had too many problems.”

Prompt: “Tell me a joke.”

The AI model, having seen the structure and style of the jokes in the examples, might generate a joke like this:

“Why don’t we allow secret agents to play hide and seek?” “Because good agents are always hard to find.”

In this case, the AI model uses the examples (the “shots”) to understand the task better and generate a response that fits the pattern established by the examples.

Chain-of-Thought Prompting

Chain-of-thought (CoT) prompting enables complex reasoning capabilities through intermediate reasoning steps. This technique can be particularly useful for complex tasks that require multiple steps or for users who prefer a more conversational interaction. The only pitfall is that it is time consuming and for especially arduous problem it may need some trials & error to arrive at the expected outcomes.

For example, you want an AI system to help you plan a birthday party. Instead of providing all the details at once, you may choose to engage in a back-and-forth conversation with the model, gradually building on previous responses. Below sample for reference:

User: “I want to plan a birthday party.”

AI: “That sounds exciting! Who is the party for and how old are they turning?”

User: “It’s for my 10-year-old daughter.”

AI: “Great! What are some of her interests? We can use those to decide on a theme for the party.”

User: “She loves unicorns and rainbows.”

AI: “Perfect! We could have a unicorn and rainbow-themed party. We can have decorations in bright colors, a unicorn-shaped cake, and games related to unicorns and rainbows. What do you think?”

In this case, each prompt from the user is based on the AI’s previous response. The AI model maintains a chain of thought throughout the conversation, building on the information provided in each exchange.

Role-Based Prompting

Role-based prompting is a technique used in prompt engineering to encourage the AI to approach a question or problem by assuming a specific role, character, or viewpoint. This technique can be particularly useful for creative writing, role-play scenarios, or any situation where you want the AI to respond in a specific role.

For example, you want AI to help you write a dialogue for a play. You can use role-based prompting by assigning roles to the AI and yourself. Below sample for reference:

User: “You are a detective, and I am a witness. Start the conversation.”

AI: “Good evening. I’m Detective Smith. I understand that you witnessed the incident this morning. Can you please describe what you saw?”

User: “I saw a man in a red jacket running away from the scene.”

AI: “Thank you for the information. Can you recall any distinctive features about this man? For example, his height, age, or any other details?”

In this case, the AI model is taking on the role of a detective and the user is taking on the role of a witness. The prompts from the user are based on the roles assigned.

Self-Consistency

Self-Consistency is a technique that aims to replace the naive greedy decoding used in chain-of-thought prompting. The idea is to sample multiple, diverse reasoning paths through few-shot/CoT, and use the generations to select the most consistent answer. This is particularly useful when there is a need to maintain a consistent logical narrative like in Q&A, Data Augmentation etc.

For example, if you ask AI to generate a review for a restaurant called “The Spicy Grill” based on the following prompts:

1. The Spicy Grill is a restaurant known for its exceptionally spicy dishes.
2. The restaurant has a cozy and warm atmosphere, making it a great place for family dinners.
3. The staff at The Spicy Grill are friendly and attentive.

A self-consistent response would be:

“The Spicy Grill certainly lives up to its name, offering a menu filled with exceptionally spicy dishes that tantalize the taste buds. The cozy and warm atmosphere of the restaurant makes it a perfect spot for family dinners. What makes the dining experience even more enjoyable is the friendly and attentive staff who ensure you have a wonderful time.”

In this response, all the information provided in the prompts is consistently reflected. The spiciness of the food, the cozy atmosphere, and the friendly staff are all mentioned, making the response self-consistent with the prompts.

Generated Knowledge Prompting

Generated Knowledge Prompting is a technique where the AI model generates potentially useful information about a given question/prompt before generating a final response.

For example, you’re trying to make a classic Italian dish, spaghetti carbonara. You know the basic ingredients: spaghetti, eggs, cheese, and pancetta. But you’re not sure about the exact steps or proportions.

Here’s where Generated Knowledge Prompting comes in. You ask, “How do I make spaghetti carbonara?” and AI generates a response based on its knowledge:

This is an example of Generated Knowledge Prompting, where AI generates a detailed response based on the prompt given by the user. The knowledge is generated on-the-fly, tailored to the user’s specific question. It’s like having a personal assistant who can provide detailed instructions or information on a wide range of topics.

Both Zero Shot and Generated Knowledge Prompting techniques rely on the AI model’s pre-training, but they differ in how they use that training. Zero-shot prompting relies on the model’s ability to generalize from its training to new tasks, while generated knowledge prompting relies on the model’s internal knowledge to generate responses.

Tree of Thoughts (ToT)

Tree of Thoughts (ToT) is a framework that encourages exploration over thoughts that serve as intermediate steps for general problem solving with language models.

For example, imagine you’re planning a vacation. In a Tree of Thoughts prompting, the conversation might look like this:

User: I want to plan a vacation.

AI: Great! There are so many aspects to consider. Would you like suggestions on destinations, budgeting, activities, or something else?

User: I’m thinking about going to Paris.

AI: Paris is a wonderful choice! You can explore various aspects like sightseeing, food, accommodation, local transport, etc. Which one would you like to know more about?

User: I would love to know about sightseeing and food.

AI: For sightseeing, the Eiffel Tower, Louvre Museum, and Notre-Dame Cathedral are popular choices. As for food, you must try the croissants, baguettes, and crêpes.

In this case, the assistant opens up multiple branches of conversation at each step, forming a tree-like structure of thoughts. This allows for a more exploratory and comprehensive conversation.

Note: It's important to note that its very similar to Chain of Thought and that the line of questions on whether it's one or the other is primarily driven by the AI in response to the user's needs. So, if you need more exploratory form of discussion then opening up the prompt with multiple options is beneficial.

Retrieval Augmented Generation (RAG)

Retrieval Augmented Generation (RAG) is a method that combines an information retrieval component with a text generator model. This might not be applicable in the general use and is more targeted at Enterprise Data systems.

For example, you want to travel to New York from Delhi and put in the prompt "What are my options to reach New York from Delhi?” in the context of an airline.

In a traditional model, the assistant might generate a response based on its pre-existing knowledge, which could look something like this:

“There are several airlines that operate flights from Delhi to New York. These include Air India, American Airlines, and United Airlines. The flight duration is approximately 15-20 hours, depending on the specific flight.”

However, in a Retrieval Augmented Generation model, the assistant first retrieves relevant information from a large database or corpus of documents before generating a response. This could include flight schedules, prices, and other details from various airlines. The assistant then uses this retrieved information to generate a more informed and accurate response. Below sample for reference:

"Based on the latest information, here are some options for your journey from Delhi to New York:

1. Srilankan Airlines: This route involves two stops - Colombo and Paris. The total duration is around 37 hours and the ticket price starts from ?46,2691.
2. Gulf Air: This also involves two stops - Bahrain and Paris. The total duration is around 50 hours and 50 minutes, and the ticket price starts from ?46,6711.
3. Cathay Pacific: This flight has one stop at Hong Kong. The total duration is around 24 hours and 5 minutes, and the ticket price starts from ?48,8261.
4. Qatar Airways: This flight has one stop at Doha. The total duration is around 31 hours and 55 minutes, and the ticket price starts from ?54,4291.

Please note that prices and schedules can vary, so it’s always a good idea to check the latest information before booking your flight."

In this case, the assistant’s response is augmented with the retrieved information, providing a more comprehensive and accurate answer.

Best Practices for Writing Effective Prompts

When creating prompts, it's essential to follow certain guidelines to ensure clarity and effectiveness.

1. Start with Clear Instructions: Always begin your prompt with explicit instructions. Clearly state what you want the model to do. For instance, if you're asking for a summary, specify that in the prompt.

Example: "Summarize the following article about climate change."

2. Be Specific: Avoid vague or ambiguous prompts. Instead, provide specific details or context to guide the model's response.

Example: Instead of "Write about cars," use "Describe the environmental impact of electric cars."

3. Precision Matters: Precision is crucial. Use concise language to convey your intent accurately.

Example: Instead of "Discuss various programming languages," try "Compare Python, Java, and C++ in terms of performance and ease of use."

4. Focus on What to Do, Not What Not to Do: Frame your prompt positively by emphasizing desired actions rather than forbidden ones.

Example: Instead of "Avoid using jargon," say "Use simple language suitable for a non-technical audience."

5. Provide Examples: Include illustrative examples to demonstrate the expected output. This helps the model understand the desired format.

Example: "Write a brief summary of the following paragraph: [insert paragraph here]."

Conclusion

Prompt Engineering is a powerful tool that can significantly enhance the capabilities of AI models. It is simultaneously both an art and a science. It is a rapidly evolving field with immense potential for future growth. As we continue to explore and understand the intricacies of prompt engineering, we can expect to see even more innovative applications and advancements in the future.

Finally, here’s a thought-provoking question to ponder: How might prompt engineering shape the future of human-AI interaction?

Next Stop: Navigating Ethical AI Landscape

I am writing a series to explore Gen AI from a beginner's point of view. Stay tuned for my next article, where we’ll review the Ethical Considerations in AI.

From identifying algorithmic biases to ensuring transparency, we’ll explore the rugged path that leads to ethically sound AI solutions.

The summit of responsible AI awaits!