In our previous article we gave insights about how SRI-CHATPED will look like and how it is tied to the PEDMaker application, which is a personalized energy companion that analyzes energy consumption/production data and provides easy-to-follow recommendations.

Now the project is nearly finished and we are thilled to present what the SRI-CHATPED chatbot can do and how the development went.

The SUSTAIN Eurocluster was hosting a networking event on Friday, November 22, 2024 at CERTH in Thessaloniki, Greece. The SUSTAIN project has received funding from the European Union's Joint Cluster Initiatives (EUROCLUSTERS) under grant agreement number 101074311. Marian Lux from LuxActive presented the SRI-CHATPED project and attended a panel discussion about >>Overcoming Challenges in Smart Building Innovation<< where he gave insights about challenges in developing smart building solutions, scaling from pilot to wide spread solutions, challenges when securing funding and the communication of ROIs to investors.

The SRI-CHATPED project aims to integrate an advanced chatbot into the existing PEDMaker application. The main objectives are to provide multimodal output, answer questions about Smart Readiness Indicators (SRI) and energy consumption/production, and offer recommendations based on live data from PEDMaker. The project solves two problems: helping prosumers save energy and educating them about SRI and how to increase their rating. To achieve these goals, the project incorporates Large Language Models (LLMs) and cutting-edge technologies such as agents, workflows, Retrieval-Augmented Generation (RAG), and Semantic Layers.

The project encountered several technical or operational challenges that required mitigation. The main challenges were in three areas: Prompt engineering, RAG on database, and history management.

• In prompt engineering, researching and testing different techniques was necessary as small changes had a significant impact. Reverse engineering state-of-the-art frameworks and altering their prompts was also done.
• In RAG on database, a semantic layer was used as a secure intermediate layer to communicate with the project's own data from a database. A sub-agent was also created to communicate with the main agent.
• In history management, existing frameworks were found to be inadequate for small models as they forgot information too quickly. To address this, a sliding window history approach was used, where relevant parts of the history were extracted and added to the query to help understand context and answer questions.

We anticipated two main risks at the beginning: hardware requirements for fine-tuning and existing solutions that could cover the planned output. To address the first risk, we used the semantic layer and RAG, which eliminated the need for fine-tuning a model. Additionally, our multi-agent architecture did not require large models.

The next steps involve scaling and further developing several areas. These include enhancing the User Interface (UI) by adding more multimodal output options such as different graph types and integrating widgets into existing applications. A framework or service will be built that can be used across various domains, incorporating a semantic layer and RAG technology.

The SUSTAIN Eurocluster and SRI-CHATPED project are working together to provide sustainable energy solutions through advanced technologies such as Large Language Models and Retrieval-Augmented Generation. The project aims to educate users about sustainable energy and the Smart Readiness Indicator (SRI) framework, while also improving the PEDMaker application. By overcoming technical challenges and mitigating risks, the project is poised for future growth and development in the field of sustainable energy.