iTop implementation for Darnytsia

Customer’s need / customer’s problem: 

Facilitate user onboarding, system adoption and e-learning process 

Lower end-user onboarding cost  

Implement modern AI toolset with minimal efforts for it’s fine tuning and initial setup 

Smartly concatenate product information with product features description to create a tool which can both facilitate the use of existing features and promote the adoption of the new. 

Enable seamless integration into the new environments. 

Consume source documentation data from external storage (Confluence) considering complex weighted data structures. 

Analysis and brainstorming challenges: 

The team performed an extensive study of available instruments and approaches to AI documentation-based chatbots development. On-prem tools were excluded due to the cost of operational expenses and focus turned to cloud-based solutions.  

The primary candidate toolset was AI Assistant from ChatGPT, available via beta version yet already demonstrating quality output and commercial capabilities with GPT-3.5 model. 

The second choice was Azure OpenAI assistant which actually is just an interface to ChatGPT based on Microsoft Azure infrastructure. 

Key competitors were among more capable GPTs which used GPT-4 as a core, yet limited as they did not provide customization capabilities required for this solution. What was also noticed, OpenAI vendor capped the performance of older products and models in favor of actual versions (which becomes challenging when you seek to build a cost-effective product). 

After series of attempts we have chosen OpenAI AI Assistant, setting a bet for OpenAI to release it from Beta with GPT-4 engine by the time we finalize the solution itself. 

In the process of tailoring the final model we have conducted an enormous number of research and testing to optimize the output and quality of responses. 

Most challenging stage of this project which crafts the core of the solution is a prompt engineering process, this consumes 70% of the total efforts. It takes pacience, dedication and multiple non-trivial decisions to make the final model correspond to the given quality target and produce correct and optimal answers. 

Final solution

The system automatically generates the necessary data packages based on source Confluence data, in several languages and transmits it to the AI assistant, specifically configured with system prompt.  

The bot interface can be embedded in any web product by means of Microsoft Power Apps with minimal effort.  

When providing responses, bot provides a link to the documentation page with full information on the subject of the initial user query. 

Bot responds to the user respecting the language of initial query even if the of documentation was not initially available in that language, which greatly expands its use when integrating into multilingual organizations. 

Benefits for customer

The end user gets a ready-to-use tool providing answers to questions from the product documentation, as well as a tool that remembers the context of communication with this user. User can navigate to the documentation using the links provided by the bot for more details. The user receives answers in the language in which he asked the questions, which helps to increase the speed of entry into the work with the product 

Feedback

I strongly believe in the future of AI for sales support tools. AI solution provided by Treran is the very first capable iteration of intelligent tools which are supporting the evolution of MySales products. 

-- Aleksey Ivasiuk, MySales Labs