:fellow professors

Michael Granitzer

Fellow Professor of Information Retrieval & Data Science

Michael Granitzer

Research group:

The Agentic AI and Retrieval group studies how intelligent agents can search, reason with, and act on information in complex environments. We combine information retrieval, applied machine learning, and foundation models to build systems that perform multi-step web search, integrate evidence from heterogeneous sources, and coordinate with humans or other agents. A particular focus is agentic information retrieval: understanding how AI agents search differently from people and how retrieval systems should adapt to these new users. The group develops open, reproducible methods and infrastructure for reliable, efficient, transparent, and controllable AI, with web search serving as both a core research problem and a real-world testbed.

Agentic AI and Retrieval

:read more

Short bio:

Michael Granitzer is Fellow Professor of Information Retrieval & Data Science at IT:U and Professor of Data Science at the University of Passau. His research is centered on the intersection of agentic artificial intelligence, applied machine learning, and information retrieval, with web search as a major application domain and research infrastructure.

He develops intelligent systems that can retrieve, analyse, and use information in increasingly autonomous ways. Building on a long-standing research record in data science, machine.

“The joy of research lies in discovering what is really there, not what we expect to find—and then using those insights to design better solutions. Our aim is to build future information systems that expand what humans and AI can discover, understand, and do.”

Publications

Title

Year

Metadata, Structure, or Strategy? A Decomposition of RAG Context Enrichment

2026

Saber Zerhoudi, Michael Granitzer, Jelena Mitrović
Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 57-74
:read more

Authors:	Saber Zerhoudi, Michael Granitzer, Jelena Mitrović
Publication date:	09/07/2026
Journal:	Joint European Conference on Machine Learning and Knowledge Discovery in Databases
Pages:	57-74
Publisher:	Springer Nature Switzerland
Description:	Retrieval-augmented generation (RAG) systems increasingly enrich retrieved passages by attaching quality metadata, structuring them into explicit records, and adopting multi-hop retrieval strategies that accumulate evidence across steps. These changes assume that richer context yields better answers, yet existing evaluations cannot test this because they vary all three factors at once. We isolate each factor in a controlled experiment across six benchmarks, four models from three families, and five enrichment levels, totaling over 24,000 evaluated responses. The assumption does not hold. Most enrichment reduces accuracy. Models prompted to use confidence scores comply correctly yet produce worse answers, a gap between utilization and accuracy that no prior work has measured. What determines answer quality is not how much metadata the context carries but whether the model can act on it for the given …
Cited by:	1
Resources:	[PDF] from arxiv.org

:view less

The Compaction Cliff in Long-Running AI Agent Memory

2026

Saber Zerhoudi, Jelena Mitrovic, Michael Granitzer
arXiv preprint arXiv:2608.22752
:read more

Authors:	Saber Zerhoudi, Jelena Mitrovic, Michael Granitzer
Publication date:	08/24/2026
Journal:	arXiv preprint arXiv:2608.22752
Description:	A safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53% of safety rules after one compaction round and 10% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its own retention policy. Three deterministic operators implement this triage across the three context-management operations: TypeCompact rewrites items in place under per-type fidelity, TypeDecompose partitions a topic too large to compact safely, replicating in-scope safety rules across partitions, and TypeRetrieve fetches items from external storage with in-scope rules pinned ahead of relevance. On five public corpora, TypeCompact preserves 2--4 more safety rules than the strongest single-shot LLM compactor at every ratio, with 96% recall over five rounds. TypeDecompose reaches 0% locality violations against 93% under uniform partitioning. TypeRetrieve reaches 100% recall@50 against 73% for the best single-shot LLM retriever. On three downstream behavioral benchmarks, we outperform the production Sonnet compactor on medical compliance (paired McNemar on preservation, ), the full-policy and hierarchical baselines on retail task pass rate (, ), and the hierarchical compaction on the airline domain (). We release AgentArtifactCorpus …
Cited by:	1
Resources:	[PDF] from arxiv.org

:view less

:all publications

:back to overview