Introduction 

Generative Artificial Intelligence (AI) has transformed the digital landscape through “intent-based outcome specification,” a paradigm where users describe desired outcomes via detailed prompts instead of traditional commands. This course – targeting a non-developer audience – delved into the foundational principles of Generative AI and Large Language Models (LLMs), focusing on their core mechanisms and capabilities. Students learned and practiced effective prompting techniques, essential for navigating this powerful yet complex method. The course included the analysis and discussion of recent research on prompt engineering, keeping students abreast of the latest developments. 

The structure of the course balanced theoretical understanding and practical application, with 30% dedicated to traditional lectures and 70% to hands-on workshops and collaborative group projects. Practical exercises using models like GPT allowed students to apply their theoretical knowledge in real-world scenarios. Group projects focused on specific application domains – including music, literature and cuisine – leading to presentations, peer reviews, and instructor feedback. 

The course – comprising 20 hours of direct instruction – was conducted at the Fondazione Bruno Kessler (FBK) campuses in Povo. Instructors Antonio Bucchiarone and Nadia Mana guided the learning journey. Additionally, Carlotta Castelluccio from Microsoft conducted a seminar on Responsible AI, emphasizing ethical considerations in AI applications. 

In this first part of the blog series, we are going to present the methodological framework and the tools used throughout the course. In the second part, we are going to cover the student projects’ main outcomes and key takeaways. 

The Card Model Template and the Flow of Cards 

One of the primary goals of the course was to provide a clear and comprehensive understanding of prompt engineering. This was achieved by introducing a structured framework known as the “Card Model” to define and organize generative AI tasks. In the context of this course, a card refers to a structured format or template used to define a specific task or objective for generating content or output using generative AI techniques. 

The Card Model serves as a conceptual framework that outlines the structure, components, and relationships involved in generating content or output using generative AI techniques. It provides a high-level abstraction of the task, capturing its essential elements and defining their interactions. Here’s a simplified model of a generative AI task: 

Objective: This is the overarching goal or purpose of the generative AI task, defining what needs to be achieved through the content generation process. 
Input: Information provided to the generative AI technique to guide the content generation process. This includes: 

Prompt: A starting point or stimulus to generate content, such as a partial sentence, a question, an instruction, or other forms of input. 
Context: Additional information or constraints that provide context for the generation task, such as background knowledge, relevant data sources, or specific requirements. 

Generative Model: The AI model responsible for generating content based on the input provided. Examples include pre-trained language models like OpenAI GPT-3.5Turbo, neural network architectures for text generation, or other generative AI systems. 
Output: The generated content produced by the generative model in response to the input, including: 

Generated Text: The actual output, which could be in the form of text, images, or other media. 
Evaluation Metrics: Criteria used to assess the quality and relevance of the generated content, including measures of coherence, relevance, fluency, and other factors depending on the specific task requirements. 

Feedback Loop: A mechanism for iteratively improving the generative AI model based on feedback from users or evaluators. This may involve refining the input prompts, adjusting model parameters, or incorporating additional training data to enhance performance. 

The Card Model helps define the key components involved in the task and their relationships, facilitating the design, execution, and evaluation of generative AI tasks in various applications. 

Cards Flow Model 

The concept of flow was also introduced in the course to provide a formal representation of the relationships between different cards composing a generative AI task. This flow model helps in visualizing and understanding the sequential and conditional transitions between different stages of a generative AI task, ensuring a structured and systematic approach to designing, executing, and evaluating generative AI processes. 

In more details, cards can be combined together to create complex workflows by defining specific transitions and dependencies between them. By linking these cards through directed edges, students can create intricate flows that mirror real-world applications. This pattern also helps students to break down complex tasks into smaller subtasks, described through detailed prompts and potentially addressed by different specialized models, generally leading to a more accurate final outcome. 

For example, a card flow might begin with a card that generates an initial story prompt. The output of this card could then flow into a card that adds contextual details, which in turn flows into another card responsible for generating the story based on the enhanced prompt. Subsequent cards could be used to evaluate the generated content, refine the prompt based on evaluation metrics, and iterate the process. 

To ensure a thorough understanding of these flows, students were asked to evaluate different paths within a flow. This involved analyzing how changes in one card could affect the overall output and exploring alternative pathways to achieve the desired outcome. Students were tasked with: 

Mapping Out Flows: Students mapped out various flows, identifying all possible paths and transitions between cards. 
Evaluating Paths: They evaluated each path to understand how different sequences and combinations of tasks impacted the final output. 
Comparing Outcomes: Students compared outcomes from different paths to determine which flow produced the most coherent, relevant, and high-quality results. 
Feedback and Iteration: They incorporated feedback into their flows, refining cards and transitions to optimize the generative process. 

By engaging in these activities, students gained hands-on experience in managing complex generative AI tasks, learning to anticipate and handle the dependencies and contingencies that arise in practical applications. This exercise not only reinforced their understanding of prompt engineering but also highlighted the importance of structured planning and iterative improvement in generative AI projects. 

The Azure AI Proxy Playground 

Students learned to interact with OpenAI models through the GUI offered by the Azure AI Proxy Playground. The service is an open-source solution which provides a Playground-like experience to explore the Azure OpenAI chat completions using a time-bound event code with different models and parameters. It’s designed for educational scenarios (e.g., a course, a hackathon, or a workshop) where students might not have access to an Azure subscription enabled with Azure OpenAI service and/or are not familiar with the Azure ecosystem and how to provision and consume Azure AI resources. 

By leveraging this solution, we were able to provide students with a simplified lab environment, where all the complexity related to the Cloud resources provisioning and model deployments was hidden to the final user and managed through a single Azure subscription, connected to the Proxy Playground. This was particularly helpful in the context of a course whose audience was non-technical and whose focus was learning to interact with large language models through prompt engineering techniques. For the sake of the course, we provisioned a gpt-3.5 turbo instance, so all the students’ interactions via the playground happened with that specific model.   

Tool GUI and Card Model mapping 

The Playground GUI is composed of several elements. Most of them can be directly mapped with the Card Model components, ensuring consistency between the theoretical concepts and the actual experimentations.  

User prompt: free-form text field used to enter the user request to the model. It’s the prompt component of the input in the card model. 
System message: free-form text field used to enter additional information to use in responses, data sources and/or tone and style specifications. It maps with the context component of the input in the card model. 
Configuration: parameters to tune the degree of randomness of the responses. It also includes a dropdown menu to select the model to use as chat engine, what we call generative model in the card template. 
Assistant response: in the chat session the user can read the model’s response, aka the generated text component of the output in the card model. 

Summary

In this article, we covered the methodological framework and tools used in the Prompt Engineering course at Fondazione Bruno Kessler, to teach non-tech students to effectively interact with generative AI models. We explored the “Card Model” – a structured approach to define and organize generative AI tasks – and the concept of the “flow”, which further structures the relationships between tasks, aiding in the creation of complex workflows.

​Microsoft Tech Community – Latest Blogs –Read More