Qwen-Based RAG Improves Ukrainian Document Understanding

Per the arXiv paper, the authors participated in the Fifth UNLP shared task on multi-domain document understanding and submitted a retrieval-augmented pipeline. The paper states the final system uses Qwen3-Embedding-8B for retrieval, a fine-tuned Qwen3-Reranker-8B for passage reranking, and Qwen3-32B for answer selection. According to the paper, reranking improved Recall@1 from 0.6957 to 0.7935, and using the top-2 reranked passages raised answer accuracy from 0.9348 to 0.9674 on a held-out split. The authors report leaderboard scores of 0.9452 (public) and 0.9598 (private).

Per the paper, the pipeline is built around three implementation choices: contextual chunking of PDFs to preserve document structure, question-aware dense retrieval, and reranking conditioned on both the question and the answer options. The authors describe constrained answer generation using a small set of reranked passages rather than broad unconstrained decoding.

Editorial analysis: Retrieval-augmented pipelines that combine large off-the-shelf embeddings with task-specific rerankers are a common pattern for contest and production tasks where document layout and multi-page context matter. Preserving document structure during chunking and making reranking answer-aware are practical levers teams use to boost retrieval precision without inventing new modeling paradigms.

For practitioners: follow whether the approach generalizes beyond Ukrainian and the shared-task data, whether reranker fine-tuning gains persist with smaller embedding models, and how constrained generation from a small passage set compares to denser retrieval with larger top-k budgets in real-world latency budgets.