BIT.UA at BioASQ 14B: Modular Retrieval with pg_textsearch and Qdrant, and Agent-Based Answer Generation

On September 7, 2026, a paper published on arXiv showed that the BIT.UA team used modular retrieval and proxy-based answer generation mechanisms in the BioASQ 14B challenge. The team replaced the original PyTerrier PISA index with PostgreSQL’s pg_textsearch for BM25 retrieval and used Qdrant to handle dense embedding indexes to support GPU acceleration. They also explored HyDE query extensions and Context-1 strategies, developing a new reordering器 training process with dense retrieval negative sampling. For answer generation, the team introduced an LLM-as-judge framework and a new proxy majority voting mechanism, using multi-agent debates and adaptive document retention to reach consensus answers, and participated in the fragment generation subtask for the first time. The system achieved competitive results in all batches.