Dorota Jasińska
Marcin Dobosz
See our recent webinar on mastering generative AI prepared by Neontri’s Technology Director, Marcin Dobosz. Marcin is a GenAI expert with rich experience in implementing GenAI solutions for enterprises. In this webinar, he talked about how to build a GenAI assistant for enterprise excellence.
Agenda:
– Neontri Assistant
– Other possible use cases
– Data processing pipelines
– Retrieval augmented generation (RAG)
– Vector database and data similarity
– Flexibility of solutions
– Closing remarks
Overview
The solution shown in the webinar is an internal Neontri’s GenAI assistant, a tool that helps process a lot of data. As Neontri works in the open banking and big data sector, such a solution was a natural use of GenAI capabilities to help organize and quickly process relevant data concerning our clients.
Neontri Assistant
The assistant implemented at Neontri works similarly to GPT, but in this case, the tool knows its role in the company. Its purpose is to help the sales and marketing department in finding necessary information about our clients without the need to involve people in the process of getting such information.
The tool has all the data about the clients and can answer questions about the projects developed by Neontri and, for example, the frameworks used in the development.
Other use cases
This solution can be implemented in different businesses and contribute to data enrichment, especially in banking transactions, etc. It can also be used to create product descriptions based on its in-store description and picture. This way, it’s possible to create a data pipeline that will include this information, process it, and create a complex description that will be easily searchable.
Another example includes giving suggestions, such as recommending which shoes to choose for a particular occasion. It will find you the product that matches the event and your outfit.
The assistant can help in finding out how customer data is being processed in a specific environment and how it’s used by a company. Such queries about terms and services can be handled by the assistant instead of a customer service agent.
What’s worth mentioning, the assistant doesn’t have to be a chat, it can be represented by another text-generating tool.
Data processing pipelines
The key data processing pipelines discussed in the solution implemented by Neontri include a loader and a chatbot. The loader pipeline handles all data stored in a source that, in this case, is Google Drive. The tool can monitor any changes to files, select relevant file types, and process the data using a toolkit (like Apache Tika). The data is then extracted, normalized, and split into chunks to be processed further.
The chatbot pipeline’s job is to provide the user with an answer to a query.
First, it analyses the query and translates it into structured data that can be easily searched in the source. It looks for relevant documents that would contain the answer to the user’s question and uses algorithms to find the best ones. Finally, it uses a language model to combine the chat history, available documents, and the user’s query to produce a response.
The pipelines also contain gateways that are used to integrate the client with the solution. The solution also includes topics that help pass data through particular steps and resources that are external systems used by the tool.
Neontri’s resources allow for scaling the tool for the client’s needs. As an example, Google Drive was presented as the source of data, and everything that happens in the Google Workspace can be processed by the tool. It’s informed about all new relevant files and folders and gets this data through the pipeline. The tool uses the data to structure it in a way that can be used to produce a high-quality response to a user’s query.
Retrieval augmented generation (RAG)
RAG allows for improving the quality of generated responses. Depending on how the language model understands the text, the assistant can create meaningful and helpful answers to queries. It compares information, looks for similarities in data vectors, and uses them to find the right answer.
Vector database and data similarity
Vector data are pieces of information stored as vectors in a database that are used to search and compare vectors. They enable fast searching for similarities used by content retrieval systems. The database represents a very big set of data, and it’s possible to run language models and our solutions on big data, like banking transactions. Vectors help select the right part of data to create an answer to the query.
Flexibility of solutions
The AI assistant can work on different language models and engines. You can connect anything that has data in existing applications and infrastructure, and it’s possible to extract the proper information for the tool. This solution can be easily implemented. There’s no need to use APIs to provide data to the solution.
Closing remarks
At Neontri, we know every enterprise needs a dedicated solution. GenAI can be used to implement those solutions while taking into consideration the requirements, infrastructure, possible constraints, and, for example, compliance issues.