A new way to shrink search AI's memory without losing its smarts.

Researchers have developed a training technique called token pooling-aware training for multi-vector embedding models like ColBERT. These models use many small vectors to represent different parts of a document, making search more accurate but also very storage-intensive. The new method teaches the AI to compress these vectors intelligently, merging similar ones while preserving the most useful information for retrieval.

For engineers building search systems, this means getting nearly the same quality results while using up to four times less storage. In tests, a model trained with this method retained 97.8% of its original retrieval accuracy while cutting document vectors by 75%. This breakthrough makes powerful, precise AI-powered search feasible for larger datasets where storage costs would otherwise be prohibitive.