Harnessing AI to Strengthen Research Impact: Initial Milestone from EDRC

Thorin Daniel, Michael Fell, Sarah Higginson, Georgia Panagiotidou, and Lirong Liu.

At centres like the EDRC and CREDS, we generate a wealth of high-quality research, evidence and recommendations. But a challenge remains: how do we ensure that these outputs meaningfully inform policy and practice? Too often, excellent insights don’t make it into consultation responses or other impact-relevant activities, simply because of time pressures or gaps in knowledge transfer across teams. Off-the-shelf AI tools (such as ChatGPT) can’t fully solve this problem. While powerful, they can’t be restricted to use only EDRC/CREDS content, making them unreliable for producing policy-facing outputs rooted in our own research.

To address this, we’re developing a database-driven solution that combines existing large language model (LLM) architectures with retrieval augmented generation (RAG).

In practice, this means taking EDRC and CREDS research outputs and converting them into a searchable database, allowing an LLM to access and draw on our research directly and accurately when generating responses.

This approach has two key applications:

• Manual use: Researchers can query the system directly to learn what our database says about specific issues.
• Automated use: The system can analyse government consultations and highlight the most relevant EDRC/CREDS materials for each question.

Together, these applications will make it easier for researchers and leadership teams to spot and respond to impact opportunities with well-prepared, evidence-based inputs. It also has the potential to extend access to the wider research community in the future.

In the short term, this project will increase the chances that EDRC and CREDS research outputs actively shape policy and practice, thus helping to ensure decisions are more evidence-informed and effective.

In the longer term, this work could serve as a model for other research centres seeking to maximise the impact of publicly funded research. We look forward to sharing more updates as development progresses and as we begin to test the system with real consultations.

Part of the questions that this project also wants to examine is to what extent can the use a RAG methodology not only support research on environmental sustainability but can also be more sustainable in itself. Accordingly, the project will observe and reflect throughout the development process to identify decision-making moments that may contribute to increased computational and subsequently resource use such as for instance, choices on data scope, chunking size or foundation model.

We are currently building bespoke repositories of content from both EDRC/CREDS outputs and open consultations. This work uses existing LLM architectures, keeping development lightweight and avoiding the high computing costs typically associated with fine-tuned AI training.