PolDense: New Polish AI models developed for information retrieval

The AI Lab at the Polish Information Processing Centre (OPI-PIB) has developed PolDense, a family of specialized artificial intelligence models designed for information retrieval in the Polish language. Unlike generative models such as ChatGPT, PolDense is a series of dense retrievers intended to support Retrieval-Augmented Generation (RAG) systems, chatbots, and AI assistants by effectively searching through large volumes of unstructured data.

The PolDense family includes six models ranging from 17 million to 1 billion parameters, built on the ModernBERT architecture. Notably, the PolDense 1B model achieved the top position in the Polish Information Retrieval Benchmark (PIRB) with a score of 64.11, outperforming significantly larger multilingual models such as Llama-Embed-Nemotron-8B and BGE-Multilingual-Gemma2-9B.

According to Marek Kozłowski, head of AI Lab at OPI-PIB, these models are designed to support the current agentic revolution by ensuring large language models can access and process accurate, locally relevant data. The development is viewed as a step toward technological sovereignty for Poland, reducing dependence on foreign corporations and providing more efficient, cost-effective solutions for administration, science, and business.