Skip to content Skip to footer

Τhe MVP of our Metasearch Platform is live!

The MVP for our Metasearch Platform developed by UPB is now live! (For a quick overview of the FUTURAL Metasearch Platform click here). This prototype utilises cutting-edge AI techniques to search and interpret data from Service Providers. Currently, it supports an open-source weather data provider, with plans to integrate additional services from the FUTURAL project soon. The MVP leverages location and time-specific data, such as precipitation probabilities, UV indices, wind speeds and directions, and estimated sunshine hours, to generate accurate, human-readable weather forecasts for any specified location and time frame.

In the FUTURAL project, the Metasearch Platform aims to provide unified access to diverse data types, including data series, charts, and heat maps, from multiple Smart Solution Providers across various domains. This data is frequently utilised by domain experts, policymakers, academics, and developers to create innovative applications, conduct research, and formulate policies. However, we recognize that this data can also be immensely valuable to non-experts, local communities in rural areas, and SMEs, helping them make informed agricultural decisions based on expert data and knowledge. Our objective is to ensure that this data is easily accessible to these users through the Metasearch platform, eliminating any technological barriers.

We believe that AI, utilising techniques for summarization and natural language generation-now achieving remarkable effectiveness through the transformer-based architecture of large language models-along with semantic data classification, are ideal tools for implementing question-answering systems. We consider these features to be essential for our platform.

Our MVP prototype is a first step in achieving this by creating a new language model to interpret expert data. To accomplish this, we have developed a specialised weather model based on the well-known Llama 2 7B language model, enhanced with over 40,000 training entries. We have rigorously evaluated the accuracy of our forecasts using four language metrics and assessments from human experts, yielding impressive results.

As the FUTURAL project progresses and more services enter the prototyping phase, we are eager to develop new service-aware models and explore innovative techniques like Retrieval-Augmented Generation (RAG). Our goal is to ensure a smooth and effective interface between Metasearch platform users and the expert-generated data from our project.