Chatbot For Database Search

Client: A large company with hundreds of employees, multiple departments, and dozens of ongoing projects at any given time

Objective: Develop a smart search system for the database which contains thousands of documents with project progress logs for the purpose of decreasing the time needed to search for information and restricting access to confidential or sensitive information based on the employee access level.

Results: We have developed a smart chatbot powered by multiple language models which reduced the time employees spend looking for information by 35%.

• Multiple language models approach

Given the confidential nature of the information stored in the database, using large language models like ChatGPT is not an option due to data leaking concerns. We have implemented a different approach to natural language processing by using a collection of multiple smaller language models to process data.

• Smart data search

When a user asks a question, the system uses a retrieval language model which looks for parts of text which might contain the answer. After the texts have been located, a reranking model ranks them in order of relevance. The system can also combine text from multiple documents to provide a complete answer.

• Data access hierarchy 

The system restricts access to sensitive information based on employee access level or department.