Bridging the Context Gap: Building AI Research Agents for Sustainable Development Goals

The article explores the critical challenge of the context gap in AI research agents, emphasizing the limitations of current models in understanding and synthesizing complex, real-world information. It introduces DeepRishSearch, a single ReAct agent designed to perform deep research on the United Nations Sustainable Development Goals (SDGs), using Retrieval-Augmented Generation (RAG) to bridge this gap. The agent operates through a structured workflow inspired by a moderated conference panel. A moderator refines user prompts by identifying keywords and mapping them to one or more of the 17 SDGs using a fine-tuned DistilBERT model. This ensures the agent focuses on relevant topics and avoids ambiguity. The refined questions are then passed to a panel of specialized "experts": a Scholar (using Semantic Scholar for academic research), a Journalist (using Tavily for recent news), and an Analyst (leveraging Our World in Data for quantitative projections and trend analysis). Each expert retrieves and augments information from their respective domains. The Scholar provides peer-reviewed insights, the Journalist delivers up-to-date developments, and the Analyst offers data-driven forecasts, including projections of SDG progress over the next five years. The agent synthesizes these inputs into a coherent report, using a small LLM (GPT-4o-mini) to summarize findings, assess trends, and assign a progress score from 0 to 10. The article demonstrates the agent’s effectiveness by testing it on Indonesia’s progress toward Zero Hunger. While a baseline model like Gemini 2.5 Pro scores Indonesia at 6/10, the DeepRishSearch agent, enriched with RAG data, assigns a more cautious 4/10. This reflects a deeper understanding of persistent challenges—such as high child stunting and wasting rates, moderate Global Hunger Index scores, and stagnant progress—despite government and international efforts. A key insight is the importance of context. Humans possess a lifetime of experiential and intuitive understanding, enabling abductive reasoning and decision-making under uncertainty. AI agents, by contrast, operate within constrained, programmatic contexts defined by their immediate inputs and tools. The agent’s strength lies not in replacing human judgment, but in augmenting it—rapidly processing vast, diverse data sources to surface actionable insights. The system is built as a public Streamlit app with open-source code, allowing users to test its capabilities. It can be adapted to other domains, such as HR (generating tailored interview questions), sales (simulating client queries), or policy analysis (synthesizing complex development data). The project underscores a broader principle: the most effective AI agents are not isolated models, but systems that integrate multiple sources, continuously learn from human feedback, and operate within a feedback loop that refines their knowledge. By closing the context gap through structured data retrieval and synthesis, such agents become powerful tools for researchers, policymakers, and analysts—accelerating discovery without replacing human expertise.