IBM has several NLP and ML powered services like Watson Assistant which enables businesses to build conversational interfaces and Watson Discovery which allows businesses to extract insights from unstructured data by understanding the meaning of data. While NLU capabilities increase the search functionalities, natural language generation however is not provided by these services out of  the box.

ChatGPT is such a hot topic now and many enterprises are showing interest to integrate their chat applications with LLMs like ChatGPT to expand from NLU to NLG but there are instances where this LLMs failed to generate correct responses. Since it's an AI model, it tries to be as accurate as it can but sometimes the responses are inaccurate. 

Here's an example - 

And there are several limitations to use ChatGPT directly in enterprise context, some such reasons include:

• ChatGPT is not hyper-personalised and brands need to be hyper-personalised to succeed in today's digital experiences world.
• Ethical AI: Bias is a sizeable issue with ChatGPT since it's trained on internet scale data which contains inherent human biases. 

How to expand enterprise chat capabilities by leveraging NLG? 

One way to expand enterprise chat capabilities to NLG is by using GPT based models tuned with enterprise data by connecting to enterprise knowledge bases like watson discovery and elastic search. To extend chat capabilities with watson assistant to support natural language generation but only limited to specific data and entrprise specific knowledge, Neuralseek can be used. Using neuralseek, NLG capabilities can be added into applications by limiting it to corporate knowledge. 

IBM Watson Assistant can extend conversational AI capabilities with NeuralSeek today. Clients can leverage queries asked in Watson Assistant, pull those out and use them to retrieve content via Watson Discovery. It then employs generative pretrained transformer technology to generate a response based on the retrieved content, the query and full context of the conversation. 

Neuralseek - https://neuralseek.com/documentation

Neuralseek offers a service for automatically training GPT (Generative Pre-trained Transformer) - based large language models (LLMs) for companies, which can be integrated with Watson Discovery, for building AI-powered search and analysis applications. Neuralseek is designed to mimic the human thought process of understanding language and provide relevant results based on the user's intent. Simply put, Neuralseek is a service which enables virtual agents to add corporate-governed natural language generation. 

One of the key components of Neuralseek is the use of large language models like GPT to generate natural language responses to user queries. With Neuralseek, companies can create their own GPT-based large language model that is automatically trained based on their specific needs and data sources. This allows companies to develop a more personalized and accurate search experience for their users.

How Does Neuralseek Work?

Neuralseek uses a deep learning model trained on large amounts of data to understand the meaning of a search query. The model uses a technique called word embeddings to represent words as vectors in high-dimensional space. This allows the model to analyze the relationships between words and infer their meanings.

Neuralseek also uses a technique called attention mechanisms, which allows it to focus on specific parts of the search query that are most relevant to the user's intent. This helps the model to provide more accurate and relevant results.

Why GPT-based Language Models with Neuralseek?

1. Personalization: GPT-based language models created with Neuralseek are tailored to the specific needs of the company, providing a more personalized search experience for users.

2. Accuracy: GPT-based language models are highly accurate, providing more precise responses to user queries.

3. Efficiency: With Neuralseek, companies can create a GPT-based language model quickly and efficiently, allowing them to provide a more effective search experience for their users.

Watson Discovery, NeuralSeek and Watson  Assistant:

The following flow describes how neuralseek  can be integrated with watson discovery and watson assistant.

The following steps can be used to integrate these services - 

1. Setup Watson Discovery

Here, I setup a collection in watson discovery which contains details about products used as goodies in IBM Think conference.

2. Integrate watson discovery with Neuralseek

i. I have integrated my collection created in watson discovery with neuralseek. And once the setup is done, I search for a query like "what is a think beanie?" and neuralseek will generate a response using NLG techniques applied on data in watson discovery. 

ii. When I search some questions however, the response is not accurate. This is because neural seek thinks I am asking for a generic product description of a polo in the following example. Most part of the answer is correct only but not really relevant and is not with applied context.

iii. To address such scenarios, neuralseek offers capabilities for misinformation tolerance. Using curate option in NeuralSeek, we can fix the answers manually and it will soon start picking up for similar questions asked. Here, I went to curate option on NeuralSeek and selected the answer I want to edit. Even though you asked one question, NeuralSeek generates similar questions that can asked using the question you asked. These can be further curated to see accurate responses. 

iv. Edit the selected part to an answer you want to see so that next time it will learn and provide similar such answers. when asked. 

3. Call Neuralseek from watson assistant

i. Go to watson assistant and create a flow to call NeuralSeek from custom extensions. The response will be as follows - 

ii. Use extension inspector with watson assistant's preview to understand how the integration is working. 

iii. Once configured into a watson assistant flow, you can see the responses in the chatbot. The preview is as follows - 

Check out this video to understand functionality of Neuralseek -