When a leading global intergovernmental body sought a chatbot to deepen stakeholder engagement on their Sustainable Development Goals (SDGs), partner Accenture turned to Llama for the organization’s first large-scale, public-facing generative AI application.

The chatbot began as a proof of concept (POC) and is designed to serve groups ranging from policy experts to the general public, providing accessible information and engagement opportunities within the initiative’s voluminous assets. Once in production, it will allow business and government leaders, policymakers, educators, students, media and everyday citizens to learn about the SDGs, their implementation progress, and ways to engage and get more deeply involved.

The organization publishes a wide range of content— everything from reports and working papers to policy briefs and more. The material currently is available on its various channels and website. Many reports are hundreds of pages long and contain content that is challenging for a general audience to understand. The chatbot can quickly and efficiently provide summaries and key points of publications related to a query that are more comprehensible for a broad audience.

“Meeting the initiative’s goals by its 2030 timeline requires a whole-of-world approach and innovative thinking,” says Lan Guan, Chief AI Officer at Accenture. “Leveraging promising technologies to engage cross-sector stakeholders and support large-scale collaboration is crucial to this effort.”

Built on Llama 3.1, the chatbot operates on AWS and employs various tools and services during customization and inference stages to ensure scalability and robustness. Training involved specific content from the organization’s publications to enhance relevance in policy recommendations and summaries. The team integrated Llama Guard to recognize and handle out-of-bound questions.

The chatbot employs a retrieval-augmented generation workflow to ground responses in the initiative’s contents, enabling content summarization and queries. The LLM can generate content by consolidating information ingested from the organization’s data sources, surfacing relevant report links as part of the response, and directing the user to visit the organization's website for deeper learning.

The organization “plays a multifaceted role in supporting the goals of this initiative, including promoting awareness, reporting on progress and enabling partnerships and collaboration,” says Guan. “The chatbot enables it to effectively engage its wide range of stakeholders on relevant, up-to-date content and highlight collaboration opportunities, thus supporting its overall strategic role in promoting the work.”

In addition to performance and safety, she says, Accenture chose Llama because it can be accessed via the cloud platform the team used for the POC, and for its multilingual capability. Leveraging an open-source LLM aligns well with the client’s goal of promoting equitable access to opportunities, innovation and technology to support development for all nations.

To address infrastructure costs and some instability associated with using the largest Llama model — Llama 3.1 405B — the team made changes to the architecture to accommodate the large model while keeping costs manageable as a temporary solution. The next phase will involve evaluating the output of a smaller Llama model and exploring fine-tuning the smaller version for comparable quality and safety outputs while keeping the chatbot highly available.

The team also plans to expand the use case to include event information related to the initiative. Plans also call for expanding beyond English to support six languages (English, French, Spanish, Chinese, Arabic, and Russian).

Further ahead, as Llama evolves and expands, Guan says the team looks forward to building a multimodal chatbot to provide the client’s stakeholders with an effective way of interacting with a wide variety of non-text-based content, including videos, graphics, and structured data.