LANGUAGE MODEL
Large language models are capable of comprehending context and semantics, allowing them to generate coherent and contextually appropriate responses.
DATA SOURCES
The model can be trained with data from documents, websites and use live data in applications such as messenger apps, CRMs and finance systems.
USER INTERFACE
Users access the chatbot via a website widget, through a connection to a messenger app such as WhatsApp or via an API integration with a compatible app.
CHATGPT WITH CUSTOM DATA
Our chatbots use ChatGPT with a custom dataset to provide extremely accurate and human-like responses to questions and requests.
ChatGPT is built on a deep learning architecture called a transformer neural network. It uses a variant of the transformer model, which is a type of self-attention mechanism that allows the model to process and understand sequences of words effectively.
TRAINING OF THE LANGUAGE MODEL
During the training process, a large dataset of text is used to train the model. The dataset consists of input-output pairs, where the model learns to predict the next word or phrase given the preceding context. This process is known as unsupervised learning, as the model doesn’t receive explicit labels for its predictions.
The training objective is to maximise the likelihood of the correct next word or phrase in each context. This is achieved through a technique called “maximum likelihood estimation,” where the model adjusts its internal parameters to improve its predictions over time.
HUMAN LIKE GENERATIVE RESPONSES
Once the training is complete, the model can generate text by taking an input prompt and predicting the most probable next words based on the context provided. The generation process involves sampling from a probability distribution over the vocabulary to produce coherent and contextually relevant responses.
To enhance the quality of responses, techniques like “beam search” can be employed. Beam search explores multiple likely word sequences during generation, considering different possibilities and selecting the most promising responses based on a scoring mechanism.
SECURITY AND DATA PRIVACY
Training a chatbot with your own company data offers a higher level of safety against threats such as indirect prompt injection and data poisoning compared to using external or publicly available datasets. Indirect prompt injection and data poisoning involves the alteration of a website with malicious content intended to manipulate the behavior of AI models. By using proprietary company data, which is carefully curated and controlled, the risk of such hidden text injections is eliminated.
Your data is stored on our secure server, or you can host the chatbot in your own data centre with our on-premise chatbot solution.
