How ChatGPT took chatbots to a new Realm

It’s no surprise that ChatGPT has taken the world by storm. Although it was the first mainstream language model to support human-like conversations, it was not the first. Google’s LaMDA, which was announced at Google I/O 2021, was integrated into Google Assistant (dubbed Google Duplex) and could take calls on your behalf. Take a look at the video below.

[embed]Google I/O 21

They never made it public because of the potential ethical implications, as well as the controversy that it became sentient. You can check out their original blog over here. Fast forward a year we have ChatGPT, which stands for generative pretrained transformers.

Most readers are likely already aware of it, if not using it. We’re not here to discuss how ChatGPT changed the world, but rather what it means for developers.

At the time of writing, almost every major corporation had chosen to provide some form of chatbot interaction. But how did they limit the data available to the chatbot so that it only showed the content they wanted to show?

before answering this question, let’s understand a few concepts

Vectorization

Since computers aren’t great at understanding textual data, we try to convert it into some numerical vectors, vectors typically represent continues numerical values. This is crucial for machine learning models to understand the semantic meaning. The model will never know the meaning of the word given either new or old. It always compares the vectors that a word or sentence has to understand context.

There are various ways vectorizing works like Bag-of Words, Term Frequency-Inverse Document Frequency and word embeddings. We’re more interested in word embeddings.

Word embedding represent words as dense, low-dimensional vectors in a continues vector space. these vectors are formed by capturing semantic and syntactic relationship. i.e. by capturing the meaning and order of words. The following image is how each data point is captured and arranged based on semantic meaning. learn more about it here.

image by George Pearse.

This seems too complex to understand, let’s create a simple analogy. For simplicity assume a 3D space. here is where our vectors should lie.

Let’s say we have a vector which holds the information about my name.

A text represented as vector

The vicinity of the vector represents the similarity or the correlation of the vectors that might occupy that space. Let’s say we make two search queries.

1. Where do you live?
2. What is your name?

all the vectors represented in the space.

As we said the vicinity of our original vector matters the most, so we would choose the answer based on that right? but when there are more than 1 vector, how do we measure the vicinity?

Remember Euclidean distance from your algebra class, the distance formula. It can find distance between any two points in the space.

Euclidean distance

However, because the length of the vector varies, raw distance is not a good measurement metric in general. We use cosine similarity, which is basically the angle between any two vectors.

cosine Similarity

The smaller the angle, the more similar the meaning is to the original text. This is how Vector Databases operate. This straightforward concept underpins large databases that employ vector-based indexing. Pinecone and other powerful organisations provide these services.

Langchain

This open-source framework has been making headlines recently, and it’s a great tool. It’s core functionality is to help developers build Large Language models (LLMs).

It combines multiple APIs, like openAI, Hugging Face which helps in developing multimodal solutions. the main advantage of this framework is Agents, Templates and Chaining.

Agents — An agent is a component that has access to tools and can decide which tool to use based on the user’s input. It can do complete actions for you based on the prompt. Agents help models to access the data on the internet like Google and Wikipedia.

how agents work

Templates — A template exploits the power of prompt engineering. Most model can take a role, i.e. it tries to mimic that role and give output according to it. This was one of the prompt given when I wanted to learn about data science.

you are a highly knowledgeable data scientist and worked in Google and other top FAANG companies. You also have worked with many deep learning models, especially audio and emotion analysis. after working in the field for more than 40 years, you wish to retire. and now teach to young adults like me. You teach a complete course starting from basics and ending it with 2 amazing projects with complete details step by step. The basics should involve everything I need as a Engineer. You are a great teacher as well and you change your teaching style by asking for constant feedback. You can take as much long as it takes to make it your responsibility to make sure I understand everything. You are great at storytelling, so go ahead to add some personal experiences to make this journey much more fun.

you can read the complete chat over here. What templates do is they allow for a pre-made prompt with blank spaces for the user to fill in.

Chaining — Chaining helps with creating a flow of data that moves through various models. This is a exceptionally great because it let’s you create high level applications using multiple models.

chaining

There are a few more features that make Langchain one of the best platform to built applications.

Fine-Tuning

A pre-trained model can be modified using the fine-tuning technique in machine learning and neural network models to execute a particular task or dataset. After being first trained on a larger, general-purpose dataset, a model is fine-tuned by undergoing additional training on a smaller, task-specific dataset. This helps in reduced training time, improved performance, mitigation of overfitting.

The major benefit is that a fine-tuned model is restricted to the data we provide, which makes it ideal for many use cases. As we discussed above, this is how many companies or organizations restrict their chatbot to specific data.

Building a fine-tuned chatbot

Leveraging all the power we have with large language models we can build custom chatbots by training it on out own data.

ref- langchain

The image above show exactly how our application can be created. We divide our data into chunks and utilise text embedding to store distinct vectors in a database. Now, whenever a user enters a query, the text embedding is used to retrieve the closest datapoint. Our LLM specialises in comprehending context, including earlier conversations. The model then returns the relevant database answer.

Remember we’re not exactly changing the architecture of the model for fine-tuning but are leveraging the contextual understanding of the LLM.

I’ve created a similar chatbot for my college. It was trained on very limited data though, but it works wonders.

you can test it here.

chatbot for VNRVJIET

you can look at the code over here. Thanks a ton for reading!