Titulo Estágio
Agentic Retrieval-Augmented Generation for High-Stake Domains
Local do Estágio
CISUC
Enquadramento
Recent breakthroughs in Large Language Models (LLMs), including GPT-4 and LLaMA, have revolutionized natural language processing through their ability to generate human-like text and reason over complex information. Their integration into applications ranging from chatbots and assistants to code generation and summarization has opened new frontiers in AI. With the advent of multimodal capabilities (text-to-image, text-to-video, etc.), these models now support a growing number of sophisticated tasks across industries.
However, LLMs remain fundamentally limited by their dependence on static, pretraining data, leading to issues such as hallucinations, outdated information, and poor adaptability in real-time, high-stakes domains. These issues are particularly critical in domains such as healthcare, financial services, legal compliance, and asset management, where errors or outdated advice can lead to serious consequences.
Retrieval-Augmented Generation (RAG) was introduced to address some of these challenges by coupling LLMs with real-time information retrieval from trusted sources. While RAG improves factual accuracy, most implementations are linear, single-pass, and non-interactive, limiting their effectiveness for complex, multi-step tasks such as medical decision support, financial analysis, regulatory interpretation, or investment strategy planning.
The emergence of LLM-powered autonomous agents introduces a new paradigm. These agents use agentic patterns such as planning, memory, reflection, and tool use to structure workflows dynamically. Their integration with RAG—referred to as Agentic RAG—allows for adaptive, iterative reasoning and real-time interaction with structured data, tools, or external APIs.
This agentic approach is especially valuable in domains where professionals must synthesize information from multiple sources, validate facts, and reason across steps, such as software development, healthcare, finance, legal and compliance, asset management.
However, no general-purpose, domain-adaptive Agentic RAG framework currently exists to support such workflows. Existing systems are either too rigid to adapt to domain-specific reasoning or too opaque to be trusted in sensitive, high-risk decisions. There is a critical need for transparent, interactive, and verifiable Agentic RAG systems that cite their sources, adapt to evolving contexts, and enable oversight by human experts.
Objetivo
This master’s thesis aims to design, implement, and evaluate a domain-adaptive Agentic Retrieval-Augmented Generation (Agentic RAG) framework that supports multi-step, source-grounded, and interactive reasoning workflows across high-stakes fields such as software development, healthcare, finance, asset management, and legal support.
Based on this research aim, the following research objectives can be identified:
1. Analyze the requirements of high-stakes domains for contextual, traceable, and adaptive AI support
• Identify core challenges in information retrieval, synthesis, and reasoning in fields such as healthcare, finance, and legal.
• Review limitations of current LLMs, RAG systems, and existing domain-specific assistants.
2. Design a generalized Agentic RAG architecture
• Define agentic patterns (e.g., planning, tool use, source tracking) tailored to complex domain workflows.
• Integrate mechanisms for citation tracking, trust calibration, and source prioritization.
• Support modular tool usage (e.g., data lookups, API access, simulations) depending on domain needs.
3. Implement and customize Agentic RAG prototypes in selected domains
• Develop proof-of-concept systems for use cases in healthcare, finance, and legal assistance.
• Enable features such as multi-turn interaction, session memory, and grounding in structured and unstructured data.
• Facilitate expert oversight and safe intervention.
4. Evaluate system performance across multiple metrics and domains
• Assess accuracy, factual grounding, completeness, and user trust.
• Measure system adaptability to real-time changes (e.g., data updates, regulatory changes).
• Conduct expert-in-the-loop studies to assess usability and reliability.
5. Reflect on generalization, limitations, and ethical concerns
• Discuss scalability to other complex workflows.
• Explore ethical and legal considerations of deploying Agentic RAG systems in regulated industries.
• Identify future research directions for robust, transparent, and safe agentic systems.
Research Questions:
1. What are the core reasoning and information challenges in high-stakes domains that existing RAG and LLM systems fail to address?
2. How can agentic workflows improve adaptability, traceability, and trust in retrieval-augmented systems?
3. What architecture and design principles enable Agentic RAG to generalize across use cases in healthcare, finance, legal, and asset management?
4. To what extent does Agentic RAG improve factual accuracy, decision-making support, and user trust compared to static LLM or traditional RAG baselines?
5. What are the key trade-offs between autonomy, human oversight, safety, and efficiency in domain-specific Agentic RAG applications?
Plano de Trabalhos - Semestre 1
1- State of the art [Sept – Oct]
2- Problem statement, research aims and objectives [Nov]
3- Design and first implementation of the AI system [Nov – Jan]
4- Thesis proposal writing [Dec – Jan]
Plano de Trabalhos - Semestre 2
5- Improvement of the AI system [Feb – Apr]
6- Experimental Tests [Apr – May]
7- Paper writing [May – Jun]
8- Thesis writing [Jun – Jul]
Condições
The work should take place at the Centre for Informatics and Systems of the University of Coimbra (CISUC) at the Department of Informatics Engineering of the University of Coimbra.
Observações
Co-orientador: Filipe Araújo
References
[1] Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., Lee, P., Lee, Y. T., Li, Y., Lundberg, S., Nori, H., Palangi, H., Ribeiro, M. T., & Zhang, Y. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4 (No. arXiv:2303.12712). arXiv. https://doi.org/10.48550/arXiv.2303.12712
[2] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2021). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (No. arXiv:2005.11401). arXiv. https://doi.org/10.48550/arXiv.2005.11401
[3] Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., Pasunuru, R., Raileanu, R., Rozière, B., Schick, T., Dwivedi-Yu, J., Celikyilmaz, A., Grave, E., LeCun, Y., & Scialom, T. (2023). Augmented Language Models: A Survey (No. arXiv:2302.07842). arXiv. https://doi.org/10.48550/arXiv.2302.07842
[4] OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., … Zoph, B. (2024). GPT-4 Technical Report (No. arXiv:2303.08774). arXiv. https://doi.org/10.48550/arXiv.2303.08774
Significant Gravitas. (2025). AutoGPT [Python]. https://github.com/Significant-Gravitas/AutoGPT (Original work published 2023)
[5] Singh, A., Ehtesham, A., Kumar, S., & Khoei, T. T. (2025). Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG (No. arXiv:2501.09136; Version 1). arXiv. https://doi.org/10.48550/arXiv.2501.09136
[6] Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models (No. arXiv:2210.03629). arXiv. https://doi.org/10.48550/arXiv.2210.03629
[7] Yildiz, D. M. (2025, January 12). How to Build AI Agents: Insights from Anthropic. Medium. https://medium.com/@muslumyildiz17/how-to-build-ai-agents-insights-from-anthropic-25e9433853be
[8] Zhang, T., Ladhak, F., Durmus, E., Liang, P., McKeown, K., & Hashimoto, T. B. (2023). Benchmarking Large Language Models for News Summarization (No. arXiv:2301.13848). arXiv. https://doi.org/10.48550/arXiv.2301.13848
Orientador
Luís Miguel Machado Lopes Macedo
macedo@dei.uc.pt 📩