ICR — Iterative Compositional Retrieval

The Problem

Most AI retrieval systems treat the context window as a container — a fixed space to be filled with retrieved information, then processed once. This approach has a hard ceiling: the context window fills, relevance degrades, and the model cannot reason across the full scope of what it needs to know.

What ICR Does

ICR is an iterative, multi-pass retrieval protocol that treats the context window as a processor, not a container.

Rather than filling the context once and hoping for the best, ICR uses multiple focused passes — each one building on the understanding generated by the last. The system reasons its way toward an answer, using retrieval as an active part of the reasoning process rather than a passive input stage.

The result is dramatically improved accuracy and coherence on complex, multi-step tasks — especially when the relevant information is spread across a large knowledge base.

Why It Matters

As AI systems are applied to increasingly complex, real-world tasks — research, analysis, long-horizon planning — the ability to reason across large bodies of information becomes critical. ICR removes the ceiling.

Publication

ICR — Iterative Compositional Retrieval
Knoechelman, J.
View Publications →