Do text embeddings perfectly encode text?

The Gradient · 2 years ago

Researchers demonstrated that text can be recovered from embedding vectors used in RAG systems and vector databases by training a model called vec2text that iteratively generates text hypotheses to match target embeddings. The method achieved 92% exact match recovery on 32-token sequences with 50 optimization steps and a BLEU score of 97. This raises security concerns for systems storing embeddings of sensitive documents, prompting future work on building embedding models that resist inversion while remaining useful.