Task
The client is a large review aggregator that planned to introduce one more feature to the product. The product is massive and has X number of new reviews appearing on the platform every day. In order to enhance the value of the product and enhance the sales of the sub-lists, they intended to incorporate a new feature through which the customers would be able to work with the reviews without the need to actually read the reviews or even compare the quantitative data manually.
Solution
In response to this request, we built an AI chatbot – an intelligent assistant that helps analyze reviews without having to actually interact with them. The client can ask the bot anything that is related to reviews about their company. For instance:
Who are the employees mentioned in the reviews?
Who has the most positive reviews?
Who has the most negative reviews?
What is the problem stated over and over again in the negative reviews?
What is the company especially good at according to the reviews?
In addition to searching and analyzing information, the bot can also create graphs based on the question entered, the trends of positive and negative reviews, etc.
Designing
The client wanted to build an AI chatbot to help with the management of reviews on the platform. This is a unique solution for his niche as competitors do not have this and the information can be obtained in the most simplified manner. When we started working, we had no design, no platform, no user interface, etc., all this was developed together with the client.
How did we do it?
All the aggregator’s feedback is saved in a PostgreSQL database. To parse them and build the context for the chatbot, we employ large LLM language models. Why LLM? The bot’s functionality is not fixed and predetermined: the customer may ask it anything they want. That is why it was crucial to use the AI that can comprehend different languages, analyze grammar, and learn on its own to be able to answer unconventional questions. Furthermore, the bot will have to work with a vast amount of data that is constantly being updated – the structure of this generative artificial intelligence enables the processing of huge amounts of data (e.g. up to 50 million pages).
After the LLM model has received the user’s input, it determines how to solve the request:
Writing SQl code ➡️ Getting the result ➡️ Writing an LLM output based on the SQl query;
Writing python code ➡️ Getting the result ➡️ Writing an LLM output based on an SQl query.
The choice of the path is related to the type of the question and is connected with restrictions on the kind of data that can be used in the SQl language. In other words, LLM can write SQl code and, therefore, give answers to the more simple questions. For example, a question about the total number of reviews about a company, the number of times a certain name is mentioned in them, or the list of reviews made within a particular time frame.
When the questions become more complex and require the analysis of several parameters at once, the LLM model uses the second variant – python code creation. For example, it will use this path when it requires to identify who are the people mentioned in the reviews the most or only include the negative reviews.
If none of the methods is suitable for solving the client’s request, the system searches for the vector storage in OpenAI and makes an API request to it to get the result.
This system is a notable combination of open source software and paid LLM models which enables the optimal use of the bot and the funds invested in it. The system is able to assess the relevance of a bot’s response: Positive ratings are used to update the large language models used. In case the bot’s response is rated negatively, then the system activates the generation of the response using the other method available within the system.
Result
The AI chatbot, thus created, offers a means of processing millions of reviews for thousands of different companies by avoiding any direct interaction with them, such as reading or counting. The bot can look for the information on any request, analyze, and even structure it. The bot is capable of handling any level of questions, from easy to complex. The uniqueness of the entire system lies in self-learning, where the customers rate the quality of the answers given by the bot. One of the big benefits this system possesses is that it is not too expensive, and this is despite the mammoth scale of data processing.