Highlights from Joint Webinar with STELAR: Smarter Agrifood with LLMs
On 18 July 2025 at 1 PM CET, the STELAR and WiseFood projects co-hosted the webinar Empowering Agrifood Applications with LLM‑based Dataset Search and Integration, moderated by Foodscale Hub.
Lasting just over an hour, the event brought together leading voices from research and industry to explore how large language models (LLMs) can support smarter data management in the agrifood sector, enrich recipes, and help citizens make healthier, more sustainable choices. The session was tailored for agrifood researchers, data scientists, policymakers, and stakeholders in digital innovation.
Setting the Scene: Introducing the Projects
Dimitris Skoutas, Principal Researcher at Athena Research Center and coordinator of both projects, opened the session with an overview. He explained how STELAR and WiseFood address complementary aspects of the agrifood data landscape.
While STELAR focuses on building infrastructure for smart agriculture and food safety data, WiseFood translates this work into citizen-facing applications that promote healthier and more sustainable diets. A shared theme connects the two: using AI‑driven methods to improve how agrifood data is managed, shared, and applied.
Inside the STELAR Project
STELAR, Spatio‑temporal Linked Data Tools for the Agri‑food Data Space, runs until August 2025 with nine European partners. The project’s mission is to address hurdles in preparing data for machine learning – such as dataset discovery, integrating heterogeneous sources, and labelling data for training models.
To meet these needs, STELAR is developing a Knowledge Lake Management System (KLMS), grounded in FAIR (findable, accessible, interoperable, reusable) and AI-ready principles. It provides tools for dataset publishing and discovery, metadata tracking, and linking workflows. Alongside this, a toolkit supports profiling, summarisation, alignment, annotation, and augmentation.
Practical demonstrations are underway through three pilots: extracting structured insights from food safety reports, predicting crop yields with satellite data, and generating suitability maps for precision farming. Together, these pilots show how better data preparation can directly enhance AI applications.
Rethinking Dataset Discovery: A Semantic Approach
Petros Skoufis, Data Science Researcher at Athena Research Center, focused on the limitations of existing dataset catalogues. Traditional keyword searches often miss the mark, struggling to interpret user intent, handle natural language queries, or support multilingual requests.
To address this, STELAR applies LLM‑based semantic dataset search. Each uploaded dataset is analysed by an LLM, which generates three levels of metadata: a general description, possible use cases, and the dataset’s domain. These embeddings are stored in a vector database. User queries undergo the same process, enabling accurate matches that go beyond simple keywords. The result is a more inclusive, precise, and multilingual‑friendly discovery experience – making datasets accessible even to non‑expert users.
From Knowledge to Action: The WiseFood Project
Launched in January 2025, WiseFood brings together nine partners to help citizens make informed dietary choices that reduce chronic disease, obesity, malnutrition, and food waste. Its aim is to encourage behavioural change by bridging two gaps: the lack of reliable nutritional knowledge and the difficulty of translating information into daily practice.
WiseFood applications combine generative AI with a multi-actor approach involving citizens, nutritionists, and food system experts. The portfolio includes FoodScholar, which organises reliable dietary information; RecipeWrangler, which enriches recipes with nutritional and sustainability insights; and FoodChat, a conversational tool for personalised meal planning. Living labs in Ireland, Hungary, and Croatia ensure that cultural diversity guides development and testing.
From Recipes to Insights: RecipeWrangler
Konstantinos Andrikos of Infili Technologies presented RecipeWrangler, created to tackle fragmented food information. At its core is an LLM agent working with recipe knowledge graphs and the HUMMUS dataset of over half a million recipes. This structure enables advanced search, in-depth insights, and transparent comparisons.
The system evaluates nutrition and sustainability indicators, proposes ingredient substitutions, and compares recipes based on cost, preparation time, or environmental footprint. With traceable and transparent outputs, users gain trustworthy, evidence-based recommendations that make healthy and sustainable eating easier.
Engaging with the Audience: Q&A Highlights
During the discussion, participants asked about the role of humans in dataset discovery and the handling of multilingual datasets. Skoufis clarified that the current system is human-centred rather than interactive, though future versions may introduce feedback loops.
Skoutas added that WiseFood collects guidelines and food composition tables in multiple languages, with English used during development but multilingual support under consideration.
The webinar concluded with thanks to the speakers and a reminder that STELAR will wrap up in August 2025.
Conclusion
The July webinar showed how LLMs can enhance agrifood data management at both technical and citizen-facing levels – from semantic dataset discovery to personalised recipe recommendations. Together, STELAR and WiseFood are delivering practical innovations for sustainable and healthy food systems.
For more insights, follow WiseFood Newsroom and its social media pages on LinkedIn, Facebook, X, Instagram, and YouTube.