TLDR: SEER is a new framework that uses large language models and agentic RAG to embed sustainability considerations into the early stages of software development, specifically during requirements engineering. It works in three stages: identifying relevant sustainability requirements (SRs) from a taxonomy and knowledge graphs, evaluating how system requirements align with these SRs, and then optimizing conflicting requirements to ensure sustainability. Tested on diverse software projects, SEER demonstrates an effective, automated approach to building more sustainable software.
The world is increasingly focused on sustainability, with the United Nations Sustainable Development Goals pushing industries to adopt greener practices. Software development, despite its intangible nature, has a significant impact on the environment, economy, and society. However, integrating sustainability considerations into software projects has traditionally been a challenge, often addressed too late in the development cycle or through high-level guidelines that are difficult to implement.
A new framework called SEER, which stands for Sustainability Enhanced Engineering of Software Requirements, aims to change this. Developed by Mandira Roy, Novarun Deb, Nabendu Chaki, and Agostino Cortesi, SEER tackles sustainability concerns right from the earliest phase of software development: requirements engineering. This innovative approach leverages the power of large language models (LLMs) and an advanced technique called agentic Retrieval Augmented Generation (RAG) to make software more sustainable from its very foundation.
Understanding SEER’s Approach
SEER operates in three distinct stages, creating a comprehensive pipeline for sustainable software development:
1. Identifying Sustainability Requirements (SRs): The first step involves pinpointing which sustainability requirements are relevant to a particular software product. SEER does this by drawing from a general taxonomy of SRs and consulting domain-specific sustainability standards, such as government guidelines. It uses LLMs to process product documentation and transform standard documents into a knowledge graph, which then helps in extracting pertinent sustainability goals, targets, and indicators. This ensures that the identified SRs are not just generic but tailored to the project’s scope.
2. Evaluating System Requirements: Once the relevant SRs are identified, SEER assesses how well the existing system requirements—both functional (what the software does) and non-functional (how well it does it)—align with these sustainability goals. This stage involves training a specialized model to identify semantically related pairs of requirements (e.g., a functional requirement and a sustainability requirement). A multi-agent LLM system then determines the exact nature of the relationship between these pairs, classifying them as positive (supporting sustainability) or negative (conflicting with sustainability).
3. Optimizing System Requirements: For any system requirements that are found to conflict with sustainability goals, SEER proposes modifications. This optimization component refines the problematic functional and non-functional requirements to ensure they positively contribute to or at least do not hinder the identified SRs. The revised requirements are then re-evaluated to confirm their improved alignment with sustainability objectives, often with human expert validation.
The Technology Behind SEER
At its core, SEER relies heavily on the reasoning capabilities of large language models, specifically using an agentic RAG approach. This means that the LLMs are not just generating text but are acting as intelligent agents, equipped with tools to retrieve information from various knowledge bases (like the SR taxonomy and knowledge graphs) and perform multi-step reasoning to arrive at accurate conclusions. The framework has been experimented with models like Gemini 2.5 and GPT 3.5, demonstrating its flexibility and effectiveness.
Also Read:
- MEC3O: A Collaborative AI Approach to Understanding Code Performance
- Optimizing LLM Code Translation for Longer Programs with Identifier Replacement
Real-World Application and Impact
The SEER framework has been tested on four diverse software projects: smart home applications, healthcare systems, e-commerce platforms, and transport systems. These case studies represent a broad spectrum of sustainability concerns, from energy efficiency in smart homes to social equity in healthcare and environmental impact in transport. The results consistently showed that SEER could accurately identify a wide range of sustainability concerns and effectively guide the optimization of software requirements.
By integrating sustainability into the very first stages of software development, SEER offers a systematic and automated way for developers to build software that is not only functional but also environmentally, economically, socially, and technically responsible. This proactive approach is crucial for achieving global sustainability goals and fostering a more sustainable future through technology. For more details, you can refer to the full research paper.
