Apple researchers propose CLaRa framework to compress documents for AI systems

Key takeaway

• Apple researchers have developed CLaRa, a framework that compresses documents into dense vectors to improve how language models retrieve and use external knowledge.

• By training retrieval and generation components together in a unified system, CLaRa reduces the amount of text models need to process while maintaining answer quality—a significant step toward making retrieval-augmented systems more practical for deployment.

3 Key Points

Context & Analysis

FAQ

How does CLaRa compress documents?

CLaRa uses embedding-based compression and a key-preserving data synthesis framework called SCP, which is based on question-answering and paraphrase supervision, to create semantically rich and retrievable compressed vectors from longer documents.

What makes CLaRa different from other retrieval-augmented generation systems?

CLaRa performs joint end-to-end training of the reranker and generator via a single language modeling loss with gradients flowing through both modules, and theoretically aligns retrieval relevance with answer quality in a unified continuous space, addressing the disjoint optimization problem in traditional RAG.

What performance did CLaRa achieve?

CLaRa achieved state-of-the-art compression and reranking performance across multiple question-answering benchmarks, even at a text compression rate of 16, outperforming text-based fine-tuned baselines.