I believe we have all tested ChatGPT, and probably like me, you have been impressed by its remarkable capabilities. However, ChatGPT has a significant limitation: it can only answer questions and perform tasks based on the public knowledge base it was trained on.

Imagine having a chatbot based on ChatGPT that communicates effectively and truly understands the nuances of your business, sector, or even a particular topic of interest. That’s the power of a custom ChatGPT. A tailor-made chatbot allows for specialized conversations, providing the needed information and drawing from a unique database you’ve developed.

This becomes particularly beneficial in industries with specific terminologies or when you have a large database of knowledge that you want to make easily accessible and interactive. A custom ChatGPT, with its personalized and relevant responses, ensures a better user experience, effectively saving time and increasing productivity.

Let’s delve into how to build such a solution. Spoiler it does not work by putting all the content into the prompt. But there is a great alternative.

Understanding the Building Blocks of Custom ChatGPT with Retrieval Augmented Generation

The foundational technology behind ChatGPT is OpenAI’s Generative Pre-trained Transformer models (GPT). These models understand language by predicting the next word in a sentence and are trained on a diverse range of internet text. However, the GPT models, such as the GPT-3.5, have a limitation of processing 4096 tokens at a time. A token in this context is a chunk of text which can be as small as one character or as long as one word. For example, the phrase “ChatGPT is great” is four tokens long.

Another challenge with Foundation Models such as ChatGPT is that they are trained on large-scale datasets that were available at the time of their training. This means they are not aware of any data created after their training period. Also, because they’re trained on broad, general-domain datasets, they may be less effective for tasks requiring domain-specific knowledge.

How Retrieval Augmented Generation (RAG) Helps

Retrieval-Augmented Generation (RAG) is a method that combines the strength of transformer models with external knowledge to augment their understanding and applicability. Here’s a brief explanation:

To address this, RAG retrieves relevant information from an external data source and uses this information to augment the input to the foundation model. This can make the model’s responses more informed and relevant.

Data Sources

The external data can come from various sources like databases, document repositories, or APIs. To make this data compatible with the RAG approach, both the data and user queries are converted into numerical representations (embeddings) using language models.

Data Preparation as Embeddings

The embeddings, which are essentially vectors, need to be stored in a database that’s efficient at storing and searching through these high-dimensional data. This is where Azure’s Cosmos Mongo DB comes into play. It’s a vector search database specifically designed for this task.

To circumvent the token limitation and make your extensive data available to ChatGPT, we turn the data into embeddings. These are mathematical representations of your data, converting words, sentences, or documents into vectors. The advantage of using embeddings is that they capture the semantic meaning of the text, going beyond keywords to understand the context. In essence, similar information will have similar vectors, allowing us to cluster related information together and separate them from a semantically different text.

Storing the Data in Vector Databases

The embeddings, which are essentially vectors, need to be stored in a database that’s efficient at storing and searching through these high-dimensional data. This is where Azure’s Cosmos Mongo DB comes into play. It’s a vector search database specifically designed for this task.

Matching Queries to Knowledge

The RAG model compares the embeddings of user queries with those in the knowledge base to identify relevant information. The user’s original query is then augmented with context from similar documents in the knowledge base.

Input to the Foundation Model

This augmented input is sent to the foundation model, enhancing its understanding and response quality.

Updates

Importantly, the knowledge base and associated embeddings can be updated asynchronously, ensuring that the model remains up-to-date even as new information is added to the data sources.

In sum, RAG extends the utility of foundation models by incorporating external, up-to-date, domain-specific knowledge into their understanding and output.

By incorporating these components, you’ll be creating a robust custom ChatGPT that not only understands the user’s queries but also has access to your own information, giving it the ability to respond with precision and relevance.

Ready to dive into the technicalities? Stay tuned!

Building the Custom “Chat with Your Data” App in Python

Now that we’ve discussed the theory behind building a custom ChatGPT and seen some exciting real-world applications, it’s time to put our knowledge into action! In this practical segment of our guide, we’re going to demonstrate how you can build a custom ChatGPT solution using Python.

Our project will involve storing a sample PDF document in Cosmos Mongo DB and developing a chatbot capable of answering questions based on the content of this document. This practical exercise will guide you through the entire process, including turning your PDF content into embeddings, storing these embeddings in the Cosmos Mongo DB, and finally integrating it all with ChatGPT to build an interactive chatbot.

If you’re new to Python, don’t worry, we’ll be breaking down the code and explaining each step in a straightforward manner. Let’s roll up our sleeves, fire up our Python environments, and get coding! Stay tuned as we embark on this exciting hands-on journey into the world of custom chatbots.

The code is available on the GitHub repository.

First, you must understand that you will need a database to store the embeddings. It does not necessarily have to be a vector database. Still, this type of database will make your solution more performant and robust, particularly when you want to store large amounts of data.

Azure Cosmos DB for MongoDB vCore is the first MongoDB-compatible offering to feature Vector Search. With this feature, you can store, index, and query high-dimensional vector data directly in Azure Cosmos DB for MongoDB vCore, eliminating the need for data transfer to alternative platforms for vector similarity search capabilities. Here are the steps to set it up:

1. Choose Your Azure Cosmos DB Architecture: Azure Cosmos DB for MongoDB provides two types of architectures, RU-based and vCore-based. Each has its strengths and is best suited for certain types of applications. Choose the one that best fits your needs. If you’re looking to lift and shift existing MongoDB apps and run them as-is on a fully supported managed service, the vCore-based option could be the perfect fit.
2. Configure Your Vector Search: Once your database architecture is set up, you can integrate your AI-based applications, including those using OpenAI embeddings, with your data already stored in Cosmos DB.
3. Build and Deploy Your AI Application: With the Vector Search set up, you can now build your AI application that takes advantage of this feature. You can create a Go app using Azure Cosmos DB for MongoDB or deploy Azure Cosmos DB for MongoDB vCore using a Bicep template as suggested next steps.

Azure Cosmos DB for MongoDB vCore’s Vector Search feature is a game-changer for AI application development. It enables you to unlock new insights from your data, leading to more accurate and powerful applications.

Cosmos DB for Mongo DB Usage Models

Regarding Cosmos DB for Mongo DB, there are two options to choose from: Request Unit (RU) Database Account and vCore Cluster. Each option follows a different pricing model to suit diverse needs.

The Request Unit (RU) Database Account operates on a pay-per-use basis. With this model, you are billed based on the number of requests and the level of provisioned throughput consumed by your workload.

As of 27th Mai 2023, the brand new vector search function is only available for the vCore Cluster option, which is why we will use this setup for this tutorial. The vCore Cluster offers a reserved managed instance. Under this option, you are charged a fixed amount on a monthly basis, providing more predictable costs for your usage.

Once you have created your vCore instance, you must collect your connection string and make it available to your Python script. You can do this either by storing it in Azure Key Vault (which I would recommend) or by storing it locally on your computer or in the code (which I would not recommend for obvious security reasons).

Custom ChatGPT boosts efficiency, personalizes services, and improves experiences across industries. Here are some examples:

• Customer Support: Companies can use ChatGPT for 24/7 customer service. With data from manuals, FAQs, and support docs, it delivers fast, accurate answers, enhancing customer satisfaction and lessening staff workload.
• Healthcare: ChatGPT can respond to patient questions using medical texts and care guidelines. It offers data on symptoms, treatments, side effects, and preventive care, helping both healthcare providers and patients.
• Legal Sector: Law firms can use ChatGPT with legal texts, court decisions, and case studies for answering legal questions, offering case references, or explaining legal terms.
• Financial Services: Banks can use ChatGPT to extend their customer service and give customers advice based on their individual financial situation.
• E-Learning: Schools and e-learning platforms can use ChatGPT to tutor students. Using textbooks, notes, and research papers, it helps students understand complex topics, solve problems, or guide them through a course.

In short, any sector needing a large information database for queries or services can use custom ChatGPT. It enhances engagement and efficiency by offering personalized experiences.

To understand the value proposition of large generative models, it can be helpful to compare them to Bidirectional Encoder Representations from Transformers (BERT) and its variations (ROBERTA, etc.).

BERT is a powerful and widely used pre-trained language model in natural language processing (NLP). It was developed by Google and released in 2018, and it quickly became one of the most influential NLP models in the field. We can consider it the predecessor of ChatGPT.

One key difference between the two models is in how they process data. GPT’s sequential processing of input sequences, one token at a time, gives it an advantage over BERT in handling longer and more complex input sequences. This makes GPT better suited for tasks requiring more intricate outputs, such as language translation and dialogue systems. Consequently, GPT is better equipped than BERT for tasks that require generating lengthier and more complex outputs.

Generative vs. Discriminative Models

Generative models like GPT and discriminative models like BERT have fundamental differences in their approach to language processing. While GPT is a large generative language model that generates new text based on input, BERT is a discriminative model that classifies text into predefined categories. While both models have unique strengths and weaknesses, their performance varies based on the task and dataset.

BERT is particularly adept at question answering, text classification, and sentiment analysis, but it may not perform as well at generating new text. On the other hand, GPT is better suited for generating new text and capturing complex language dependencies. This makes it ideal for content generation, language translation, summarization, and question-answering.

Pretraining

Another important aspect of GPT’s training methodology is pre-training. Before being fine-tuned for specific tasks, GPT is pre-trained on a vast amount of data, learning to generate text by predicting the next word in a sentence.

This pre-training phase helps GPT learn grammar, facts about the world, and gives the model even reasoning abilities. This general language understanding serves as a strong foundation for GPT when it comes to solving specific natural language tasks later on. By leveraging this pre-training, GPT can easily adapt to new tasks with relatively less task-specific data. This process, called transfer learning, enables GPT to perform better than other models in various tasks.

Performance vs. Capabilities

Performance and capabilities are distinct factors when evaluating language models. While BERT excels in some applications, GPT’s strengths lie in its capabilities across various fields, particularly with few-shot or zero-shot learning. By fine-tuning GPT to specific tasks, its performance can be further improved and may likely outperform BERT.

Although GPT is proficient at basic NLP tasks like sentiment analysis and text classification, performance comparisons show that BERT can achieve similar or better results with less computational complexity in fundamental NLP tasks. However, the performance of GPT-4, which is yet to be seen, may likely outperform BERT in almost any discipline, even without fine-tuning.

Unmantling GPTs Value Proposition

Despite the impressive capabilities of generative language models like ChatGPT, examining their value proposition in more detail is essential. Therefore, this article aims to provide a more nuanced understanding of the value that these models can provide.

Also: Eliminating Friction: How OpenAI’s GPT Streamlines Online Experiences and Reduces the Need for Traditional Search

Large generative language models like ChatGPT offer valuable benefits to businesses through their ability to generate natural language responses similar to those produced by humans. This technology can be used in various ways, such as content generation, customer service, and marketing.

Businesses can use generative language models to produce high-quality content quickly and efficiently. For example, a news organization could use ChatGPT to generate news articles or summaries based on current events. Similarly, a company could use this technology to create product descriptions, emails, or even social media posts.

Generative language models can also be employed to provide customers with instant responses to their queries, which could be particularly useful for businesses that receive a high volume of customer inquiries or support requests. ChatGPT can be trained to provide accurate and helpful responses to frequently asked questions or to engage in more complex conversations with customers.

In marketing, generative language models can be used to create personalized content for customers by analyzing customer data to generate customized marketing messages or entire campaigns that resonate with individual customers’ preferences and interests.

ChatGPT’s ability to handle longer input sequences enables it to maintain context and understand the sentiment behind a piece of text more effectively. The use of self-attention mechanisms allows ChatGPT to focus on the most relevant parts of the input when generating its predictions, leading to more accurate results in tasks like sentiment analysis. Additionally, ChatGPT’s increased capacity allows it to learn more complex patterns and representations, resulting in improved performance across various natural language tasks.

For smaller organizations with limited data science resources, implementing AI in their processes can be a significant challenge. Developing specialized models for tasks such as summarization, classification, and translation requires substantial expertise and training data. In many organizations, these resources are not readily available, which can slow down development processes and hinder innovation.

GPT’s versatility addresses this challenge by offering a single API that can perform these tasks and more. This enables smaller organizations to benefit from AI without the need to invest in extensive data science resources. By automating and streamlining their workflows, these organizations can save time and resources, allowing them to focus on their core activities.

A lot of the versatility comes from GPT, allowing for zero-shot or few-shot predictions. Zero-shot learning is a technique where a model is able to perform a task without any explicit training examples. This is possible because GPT was pre-trained on almost the entire text available from the public internet. It allows the model to make inferences based on the patterns it has learned from the data. Few-shot learning, on the other hand, involves training a model on a small amount of data.

It’s important to note that using GPT also poses potential risks, such as biases and inaccuracies. Smaller organizations may lack the resources to address these risks and, therefore, must evaluate GPT’s performance carefully before integrating it into their processes. Nonetheless, the availability of GPT represents a significant opportunity for smaller organizations to leverage AI in their operations and remain competitive in their respective markets.

One of the major benefits of ChatGPT and Davinci is their ability to perform multiple tasks within a single request. For instance, a prompt to a GPT model that asks for a summary in five sentences and a German translation can effectively combine the tasks of summarization and translation. This multi-tasking capability streamlines the development process and simplifies complex procedures.

GPT – the Swiss Army Knife of AI

Imagine a situation where a process involves several tasks like translating customer requests, checking specific information, categorizing, and summarizing them. Traditional models would need the creation, integration, security, and maintenance of four separate models. However, a multi-purpose language model like GPT can handle all these tasks in just one request all at once.

While other models like BERT can perform tasks such as language translation and text classification, the ability of ChatGPT and Davinci to execute multiple tasks at once sets them apart. By moving some of the complexity into a prompt for a model, organizations can adapt more easily to changing requirements and become more agile.

ChatGPT and Davinci can be seen as the Swiss Army Knives of AI language models. They offer versatile and adaptable solutions for a wide range of tasks. Much like a Swiss Army Knife, these multi-purpose models provide organizations with a valuable tool that simplifies and streamlines complex procedures, making them an essential asset in today’s rapidly evolving world.

An Ongoing shift Toward AI

As generative AI technology continues to advance, an increasing number of organizations are likely to rely on these models to help simplify their complex processes. This shift can lead to improved efficiency, cost savings, and enhanced accuracy, enabling businesses to focus on their strategic objectives. However, this transition also brings potential risks and challenges, such as ensuring ethical AI usage and addressing the possibility of job displacement. Organizations must carefully consider these factors as they integrate AI into their operations.

The multi-tasking abilities of ChatGPT and Davinci offer a distinct advantage for organizations aiming to streamline intricate processes and boost efficiency. By delegating some of the process complexity to these models, businesses can adapt more rapidly to evolving requirements and improve their overall agility. Nevertheless, it is essential for organizations to assess the potential challenges, ethical considerations, and workforce implications as they incorporate AI into their operations. By doing so, they can make well-informed decisions and develop a balanced approach to harnessing the power of generative AI models, ultimately ensuring sustainable growth and responsible AI integration.

A major advantage of OpenAI, including GPT, is its ability to lower the entry barrier for organizations using AI. GPT is accessible to developers and data scientists of all skill levels, making it easier for organizations to automate activities without extensive expertise. Its capacity to generalize to new cases (zero or few-shot learning) allows users to start with OpenAI even with little or no data. This is particularly beneficial for smaller customers. They may lack resources for in-house predictive model development, as well as larger customers who can speed up their development processes using a single multi-purpose AI.

Moreover, OpenAI operates as a cloud service, eliminating the need for organizations to build and maintain their own AI infrastructure for GPT model development and hosting. Instead, they can utilize the cloud-based service provided by OpenAI, making it more convenient and cost-effective to begin using AI. This approach allows businesses to concentrate on their core competencies while leveraging GPT’s power to enhance operations and drive innovation.

The scalability of Azure OpenAI also empowers businesses to start with a proof-of-concept project and scale up as required. This approach enables organizations to experiment with AI without committing to a large initial investment. Utilizing a single model for various purposes significantly accelerates the creation of POCs. Once a solution demonstrates its value, organizations can later fine-tune the process using more specialized models.

This article has explored the unique value proposition of OpenAI’s GPT in a business context, highlighting its enhanced language capabilities (1), versatility in use (2), complexity reduction (3), and lower entry barriers for AI adoption (4). These aspects make GPT a groundbreaking development in the field of artificial intelligence, particularly within natural language processing (NLP).

GPT has demonstrated impressive performance across a wide array of applications, such as chatbots, personalized content generation, question-answering systems, and intricate data interpretation. While other NLP models can accomplish some tasks carried out by GPT, its extensive pre-training on large data sets and ability to manage various domains and tasks render it more flexible and powerful. Consequently, GPT’s potential to streamline workflows and reduce costs is indisputable.

As OpenAI continues to advance and refine its technology, we can anticipate even more innovative use cases for GPT in the future. This ongoing evolution will undoubtedly contribute to the growing significance of GPT in shaping the AI landscape and revolutionizing the way businesses harness the power of artificial intelligence.

Sentiment analysis is the process of identifying the sentiment, or emotional tone, of a piece of text. This can be useful for a wide range of applications, such as identifying customer sentiment towards a product or service, or detecting the overall sentiment of a social media post or news article.

Sentiment analysis is typically performed using natural language processing (NLP) techniques and machine learning algorithms. These tools allow computers to “understand” the meaning of text and identify the sentiment it contains. Sentiment analysis can be performed at various levels of granularity, from identifying the sentiment of an entire document to identifying the sentiment of individual words or phrases within a document.

How Sentiment Classification Works

There are many different approaches to sentiment analysis, and the specific methods used can vary depending on the specific application and the type of text being analyzed. Some common techniques for performing sentiment analysis include using machine learning algorithms to classify text as positive, negative, or neutral, and using lexicons, or lists of words with pre-defined sentiment, to identify the sentiment of individual words or phrases. In this way, it is possible to measure the emotions towards a specific topic, e.g., products, brands, political parties, services, or trends.

We can show how sentiment analysis works with a simple example:

• “This product is excellent!”
• “I don’t like this ice cream at all.”
• “Yesterday, I’ve seen a dolphin.”

While the first sentence denotes a positive sentiment, the second sentence is negative, and in the third sentence, the sentiment is neutral. A sentiment classifier can automatically label these sentences:

Text Sequence	Sentiment Label
This product is great!	POSITIVE
I wouldn’t say I like this ice cream at all.	NEGATIVE
Yesterday I saw a dolphin.	NEUTRAL

Predicting sentiment classes opens the door to more advanced statistical analysis and automated text processing.

Sentiment analysis is used in various application domains, including the following:

• Sentiment analysis can lead to more efficient customer service by prioritizing customer requests. For example, when customers complain about services or products, an algorithm can identify and prioritize these messages so that sales agents answer them first. This can increase customer satisfaction and reduce the churn rate.
• Twitter and Amazon reviews have become the first port of call for many customers today when exchanging information about products, brands, and trends or expressing their own opinions. A sentiment classifier systematically enables businesses to evaluate this information. It can collect data from social media posts and product reviews in real-time. For example, marketing managers can quickly obtain feedback on how well customers perceive campaigns and ads.
• In stock market prediction, analyze the sentiment of social media or news feeds towards stocks or brands. The sentiment is then used as an additional feature alongside price data to create better forecasting models. Some forecasting also approaches exclusively rely on sentiment.

Sentiment Analysis will find further adoption in the coming years. Especially in marketing and customer service, companies will increasingly use sentiment analysis to automate business processes and offer their customers a better customer experience.

How Sentiment Analysis Works: Feature Modelling

An essential step in the development of the Sentiment Classifier is language modeling. Before we can train a machine learning model, we need to bring the natural text into a structured format that the model can statistically assess in the training process. Various modeling techniques exist for this purpose. The two most common models are bag-of-words and n-grams.

Also: 9 Powerful Applications of OpenAI’s ChatGPT and Davinci

Bag-of-word Model

The bag-of-word model calculates probability distributions over the number of unique words. This approach converts individual words into individual features. Fill words with low predictive power, such as “the” or “a,” will be filtered out. Consider the following text sample:

“Bob likes to play basketball. But his friend Daniel prefers to play soccer. “

Through filtering of fill words, we convert his sample to:

“Bob”, “likes”, “play”, “basketball”, “friend”, “Daniel”, “play”, “soccer”.

In the next step, the algorithm converts these words into a normalized form, where each word becomes a column:

The bag-of-word model is easy to implement. However, it does not consider grammar or word order.

What is an N-gram Model?

The n-gram model considers multiple consecutive words in a text sequence and thus captures word sequence. The n stands for the number of words considered.

For example, in a 2-gram model, the sentence “Bob likes to play basketball. But his friend Daniel prefers to play soccer.” will be converted to the following model:

“Bob likes,” “likes to,” “to play,” “play basketball,” and so on. The n-gram model is often used to supplement the bag-of-word model. It is also possible to combine different n-gram models. For a 3-gram model, the text would be converted to “Bob likes to,” “likes to play,” “to play basketball,” and so on. Combining multiple n-gram models, however, can quickly increase model complexity.

Sentiment Classes and Model Training

The training of sentiment classifiers traditionally takes place in a supervised learning process. For this purpose, a training data set is used, which contains text sections with associated sentiment tendencies as prediction labels. Depending on which labels we provide and the training data, the classifier will learn to predict sentiment on a more or less fine-grained scale. Capturing neutral sentiment requires choosing an odd number of classes.

More advanced classifiers can detect different sorts of emotions and, for example, detect whether someone expresses anger, happiness, sadness, and so on. It basically comes down to which prediction labels you provide with the training data.

When the classifier is trained on a one-gram model, the classifier will learn that certain words such as “good” or “great” increase the probability that a text is associated with a positive sentiment. Consequently, when the classifier encounters these words in a new text sample, it will predict a higher probability of positive sentiment. On the other hand, the classifier will learn that words such as “hate” or “dislike” are often used to express negative opinions and thus increase the probability of negative sentiment.

Language Complications

Is sentiment analysis that simple? Well, not quite. The cases described so far were deliberately chosen to be very simple. However, human language is very complex, and many peculiarities make it more difficult in practice to identify the sentiment in a sentence or paragraph. Here are some examples:

• Inversions: “this product is not so great.”
• Typos: “I live this product!”
• Comparisons: “Product a is better than product z.”
• In a text passage, expression of pros and cons: “An advantage is that. But on the other hand…”
• Unknown vocabulary: “This product is just whuopii!”
• Missing words: “How can you not this product?”

Fortunately, there are methods to solve the complications mentioned above. I will explain more about them in one of my future articles. But for now, let’s stay with the basics and implement a simple classifier.

Venturing into the practical aspects of sentiment classification, our aim in this tutorial is to create an efficient sentiment classifier. Our focus will be on a dataset provided by Kaggle, comprising tens of thousands of tweets, each categorized as positive, neutral, or negative.

Our objective is to design a classifier capable of assigning one of these three sentiment categories to new text sequences. To this end, we will employ two distinct algorithms – Logistic Regression and Naive Bayes – as our estimators.

The tutorial culminates with a comparative analysis of the prediction performance of both models, followed by a set of test predictions. Through this hands-on approach, you will gain an understanding of the nuances of sentiment classification and its application in understanding public opinion, especially on social media platforms like Twitter.

Boost your sentiment analysis skills with our step-by-step guide, and learn to leverage machine learning tools for precise sentiment prediction.

The code is available on the GitHub repository.

Next, let’s quickly clean and preprocess the data. First, as a best practice, we will transform the sentiment labels of the train and the test data into numeric values.

In addition, we will add a column in which we store the length of the text samples.

• Accuracy is calculated as the ratio between correctly predicted observations and total observations.
• Precision is calculated as the ratio between correctly labeled values and the sum of the correctly and incorrectly labeled positive observations.
• The formula for Recall is the ratio between correctly predicted observations and the sum of falsely classified observations.
• F1-Score takes all falsely labeled observations into account. It is, therefore, useful when you have an unequal class distribution.