Shanghai AI Lab's IA-RAG Beats Standard Search on Complex Time-Based Questions by 14%

Ask a standard RAG system "where did Karl Menger work three years and two months before February 1930?" and it will almost certainly get it wrong. The system finds chunks mentioning Menger near 1930, sees the University of Vienna, and stops there. The correct answer requires working backwards through a precise time interval to land in October 1926, then finding that Menger was at the University of Amsterdam at that point. That kind of arithmetic over overlapping time windows is exactly what current retrieval systems cannot do. A new paper from researchers at East China Normal University and Shanghai AI Lab proposes IA-RAG, a framework that fixes this by treating every fact as a time interval and using formal temporal logic to navigate between them.

Why timestamps alone keep failing

The core problem is that most RAG systems, including graph-based ones like GraphRAG and LightRAG, treat time as a shallow metadata tag. A fact gets a year, maybe a month, and that's it. Real-world facts are not isolated time points; they are intervals with duration, overlap, and containment. This creates three concrete failure modes:

• Temporal boundary ambiguity: Semantic retrievers struggle to distinguish between states of the same subject across different periods. Queries like "policies before 2008" and "policies after 2008" yield nearly identical embeddings, causing retrieval of temporally irrelevant facts.
• Missing timestamps: Many real-world facts lack explicit timestamps. Current methods fail to utilize the complementary relationships between events to infer or impute these unknown temporal intervals, leaving a significant portion of the knowledge base temporally unanchored and unsearchable.
• Topological reasoning gaps: Complex queries often require inferring relationships between multiple events, such as determining if one event occurred during another or if they overlap. Existing RAG systems return disjoint chunks without modeling these interval-based dependencies, forcing LLMs to rely on brittle linguistic inference rather than structured temporal logic.

The algebra behind the fix

IA-RAG is built on Allen's Interval Algebra, a formalism introduced by James Allen in 1983. The calculus defines possible relations between time intervals and provides a composition table that can be used as a basis for reasoning about temporal descriptions of events. The key insight is that any two time intervals can only relate to each other in exactly 13 ways: one can be before, after, during, contains, overlaps, meets, starts, finishes, and so on, plus their inverses. Using this calculus, given facts can be formalized and then used for automatic reasoning. IA-RAG is the first retrieval framework to apply this formalism directly to knowledge retrieval at scale.