Make your IT Services AI-Ready

TL;DR:?

• You can make a "custom AI" by using your own data.
• 'Embedding' is a handy trick that turns the meaning of text into numbers that a computer can understand.
• You can compare these numbers to find out which texts have similar meanings.
• Using Embedding with your data is easy, won't cost much, and can make a big difference in how you can use AI today.
• You can use Embedding today to do tasks like ticket routing or a chatbot based on internal processes and knowledge.

In today's rapidly advancing digital landscape, it's crucial to identify the tangible techniques that can help us tap into the immense potential of Artificial Intelligence. Rather than focusing solely on more complex algorithms, we must place equal emphasis on the fundamental aspect often overlooked: data. Maintaining clean, well-organized data is the stepping stone to effectively harness the power of AI. This article will delve into the importance of well-curated data, and more specifically, we'll explain and explore the concept of embeddings that help translate this data into actionable insights. We will provide you with some guidance to make your company 'AI-Ready'. And finally, we will look into real use-cases for IT Services.

The Imperative of Clean Data

In the landscape of artificial intelligence, well-organized and meticulously maintained data serve as the bedrock. Just as an organized library allows for efficient information retrieval, clean data enables precise AI functioning.

If you want to have your "own AI", meaning one personalized with knowledge about your data, you have three options. These options range from the most complex, expensive, and data-intensive to the simplest, cheapest, and least data-dependent:

1. Training your own model: Developing a robust AI model from scratch requires a substantial volume of clean data. Along with the initial data requirement, training your own AI model also involves considerable expense.
2. Fine-tuning an existing model: Fine-tuning involves using an existing model, like GPT3, and training it further with specific data. In this scenario, the volume of data required decreases, but the need for clean, relevant data persists. The cost is also significantly lower than training a model from scratch.?
3. Use a base model with contextual prompting: Using a base model with contextual prompting: Clean data is still essential here, but the size of the dataset is less relevant and depends on your use case. For instance, you could use a single PDF document. The setup cost is close to zero, and the use of the model is usually very cheap.

In all three scenarios, the absence of clean and well-structured data impedes AI's ability to learn, adapt, and deliver accurate results. Therefore, nurturing the quality of data is a non-negotiable step towards harnessing the full potential of AI.

Later in this article, I will provide some best practices for managing your data. But first, let's explore how to implement contextual prompting using a technique called Embedding.

Understanding Embedding

In simple terms: Embedding is a technique to turn the meaning of a text into a numerical representation. This isn't about digitizing words, but about capturing the underlying context and semantics of the text. This numerical representation is typically called a vector, which is capable of understanding meaning, not the exact text.

Here is a quick vocabulary summary, which may clarify some confusion when researching online:

• Embedding: This is the technique used to convert meaning into numerical representation.
• Vectors, Embeddings, or Embedding Vectors: These terms refer to the resulting numerical representations.

In this article, I will use "Embedding" to talk about the technique and "Vectors" to talk about the numerical representations.

Now, let's address the question: How do we use these vectors? Assume we introduce a new word like "dog" into our current example. To discover words that carry similar meanings, we transform "dog" into a vector and perform a cosine similarity comparison. Simply put, we compare the vectors, and those that show the highest similarity typically convey the most similar meanings.

However, as I briefly mentioned earlier, the exact textual representation of a word can't be retrieved from its vector representation. This is why it's crucial to store the text, the vector, and any associated metadata together. Now, let's take a swift dive into Vector Databases and then touch on some best practices for managing your data.

Understanding Vector Databases

Vector Databases are tailored for storing and retrieving high-dimensional data, including vectors, text, and metadata.

• The Text: The original content that the vector represents, ranging from single words to whole documents.
• The Vector: The numerical representation. It holds the meaning of the text in a machine-understandable format, crucial for tasks like similarity search.
• The Metadata: Additional information related to the text or vector, like title, date, or author. This extra layer of context refines search results and ensures relevancy.

The magic of Vector Databases lies not just in storage but in their ability to manipulate and retrieve vectors efficiently. They perform similarity-based searches, identifying the closest matches amongst billions of vectors.

Here are some benefits of using Vector Databases:

• Efficient Retrieval: They enable quick and efficient retrieval of data, especially when dealing with high-dimensional vectors.
• Scalability: They can handle billions of vectors effectively.
• Improved Search: By storing vectors and metadata, they allow for nuanced search capabilities, including similarity search, providing more relevant and contextually appropriate results.
• Real-Time Analysis: They can perform real-time analysis on high-dimensional data.

For those interested in Vector Databases, open-source option like Chroma and cloud-based solution like Pinecone Pinecone are available and widely used, along with offerings from Google Cloud and other cloud providers.

Building a Strong Data Management Framework

Now that we comprehend the power of using vectors and how they're stored, we can utilize this knowledge to implement effective data management practices. The objective is to structure and maintain data in a way that is both easily understood by AI models and easy to turn into vectors. Here are some key practices:

• Explicit Partitioning: Clearly delineate between the primary data carrying the semantic meaning, and the auxiliary information known as metadata, such as titles or tags.
• Chunking: Instead of storing data as one large entity, break it into manageable, digestible segments or "chunks". These chunks can then be individually encoded as embeddings without losing any important context or meaning.
• Update Tracking: Incorporate a system to attach 'last updated' timestamps to each data chunk. This not only helps in tracking the freshness of data but also enables easy identification of data segments that need to be converted.

By adhering to these practices, you cultivate a data environment conducive to the optimal operation of your AI models.

In the upcoming section, we will immediately dive into tangible use cases in IT services.

Practical Use Cases of AI in IT Services

Let's now look at how the ideas we've just discussed can improve IT services.

Use Case 1: Ticket Routing

A valuable area of AI application is ticket routing. Rather than requiring team members to determine where a support ticket should be directed, an AI model can handle this decision. This reduces manual intervention and enhances efficiency.

To implement this, use your existing knowledge bases about ticket routing and apply the techniques we've discussed. Transform the user's request into a vector, and compare it to your Vector Database.

Moreover, you can use AI models like OpenAI 's ChatGPT to extract the most important parts of the request to aid with embedding. You can also consult with ChatGPT to verify the accuracy of your results and, if required, return a default value for manual handling.

Use Case 2: AI as an internal Knowledge Base

The same process can be applied here as well. The distinction lies in utilizing a Vector Database that contains information about your internal processes.

For instance, when an employee needs to know how to set up their work-from-home VPN, they can simply ask the AI. Your code will review the vectors of all documents, identify the most relevant ones based on the query, and adds the information retrieved when asking the AI. It can guide the employee to the appropriate document or even provide a summary of the necessary steps.

Many more use cases

I have only briefly described two use cases, but here are some other ideas:

• Crisis management: While in crisis, you can ask the AI for the next steps and actions to take based on your companies processes.
• Onboarding: One of the most important parts of the journey of your employee yet one of the least prioritized. Let the AI talk with the new employee about your culture, rituals and other processes.
• Ticket qualification: The AI will ask questions to have all the elements necessary to troubleshoot an incident or handle a request.

You can already use AI and make a proof-of-concept

AI's potential to revolutionize IT services is not a distant reality but a practical, achievable goal that can be brought to life today. By harnessing AI for tasks like ticket routing and creating an AI-driven knowledge base, we can incrementally transform our IT service delivery. As we actively refine our data handling strategies and deepen our understanding of AI, we lay the foundation for our organizations to adapt and thrive in an evolving technological landscape. The journey begins with simple proofs of concept and leads to a significant shift from traditional methods towards more streamlined, efficient, and innovative processes. I’m always interested in hearing about how you’re using AI for your IT Services, feel free to reach out to me anytime.