International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 10 Issue: 07 | July 2023
p-ISSN: 2395-0072
www.irjet.net
Designing a Generative AI QnA solution with Proprietary Enterprise Business Knowledge using Retrieval Augmented Generation (RAG) Suvoraj Biswas Solutions Architect, Ameriprise Financial, Minneapolis, Minnesota, USA ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Large Language Models from OpenAI’s
The traditional Enterprise search system depends on the regular full text search or partial text search and lists down the knowledge sources or articles based on the exact word matching. This sometimes pulls the incorrect sources of information or too much information.
ChatGPT or Google’s BARD have the capability to generate human-like responses in natural language. This capability can be used to design solutions to solve many enterprise business use cases. In this prototype solution we are trying to design an Enterprise content search solution using Generative AI. This QnA (Question and Answering) framework would be designed based on OpenAI’s APIs on top of the private business knowledge for internal stakeholders of an organization. This solution would try to leverage the summarization and embeddings generation capabilities of OpenAI’s API as well as Vector Database as part of the private knowledge repository in the solution. In the prototype solution we will measure the cost of the Q&A system based on OpenAI's offerings with different types of LLM models for a fixed knowledge dataset.
Our proposed Generative AI based solution would help the enterprise stakeholders to correctly point out the exact response or steps/process flows out of the tons of knowledge articles. The solution outlined below would also use the Large Language Model’s summarization capability to provide exact responses so that users do not need to browse through the knowledge sources to identify the information they are looking for. The process is called Retrieval Augmented Generation (RAG) where the LLM model is used to generate human readable response in the natural language while setting the context or boundary within the Enterprise business knowledge so that the LLM model doesn’t hallucinate or generate incorrect response.
Key Words: Generative AI, LLM, Embeddings, Vector Database, Pinecone, Langchain, Open AI, GPT (Generative Pre-trained Transformer), Machine Learning, Solution Architecture, Enterprise AI Knowledge framework, Retrieval Augmented Generation (RAG) framework.
That would definitely help the enterprise to save tons of business hours with a high customer satisfaction rate. In the following sections we will cover some important concepts of AI which are the basic building blocks of our proposed solution.
1. INTRODUCTION Content is an integral part for any Enterprise. The contents or business knowledge are useful for internal stakeholders who consume this knowledge about a process or workflow and complete a specific workstream. Consider the following problem statements and use cases:
1.1 Introduction to Embeddings Embedding is one of the major building blocks in our solution. Embeddings refer to the mathematical representation of a piece of text or words or graphic contents such as images or media contents (video/audio) in such a way that it becomes easier to find the closeness or relatedness of those data. E.g. consider the following three sentences -
a) Airline industry- It uses various internal and external applications for managing bookings/reservations or passengers data or fleet schedules. An internal employee like a booking agent has to have good business knowledge to serve the external customers. The agent spends a huge time figuring out the correct workflow by referring to the proprietary enterprise knowledge articles.
a) Peter loves eating cheese pizza more than anything b) ChatGpt is disrupting everything c) Dominos is offering some really cool deals For a human it is very easy to figure out that the two sentences (a) and (b) have some closeness since both are connected with pizza, however the 2nd sentence (b) has no relatedness with the rest of the sentences. If these above three sentences are plotted against a three dimensional graph it would probably look like below :
b) Financial organizations- They have built a huge knowledge and research repository based on the market research done by their analysts over time but finding the correct step or referring to the correct research is a huge pain when the information is in a case study format.
© 2023, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 382