BioASQ14b

Codebase for the BIT.UA team's (University of Aveiro) participation in the 14th edition of the BioASQ Task B challenge on biomedical question answering.

This repository is the official implementation accompanying the paper:

André Ribeiro, Rúben Garrido, Alexander Christiansen, Richard A. A. Jonker, Sérgio Matos. "BIT.UA at BioASQ 14B: Modular Retrieval with pg_textsearch and Qdrant, and Agent-Based Answer Generation." CLEF 2026 Working Notes. arXiv:2609.04999

System Overview & Approaches

The pipeline tackles the primary BioASQ Task B phases through a modular architecture:

1. Phase A — Document Retrieval

A hybrid multi-stage retrieval pipeline designed for scalable search and precise ranking over biomedical literature (PubMed):

• Lexical Search (BM25): Replaces traditional inverted-index tooling with PostgreSQL-based pg_textsearch, enabling fast BM25 scoring and dynamic index updates directly inside the database.
• Dense Retrieval: Vector embeddings served via Hugging Face Text Embeddings Inference (TEI) and indexed in Qdrant for GPU-accelerated similarity search.
• Query Expansion: Incorporates Hypothetical Document Embeddings (HyDE) to bridge vocabulary gaps between colloquial questions and medical literature, alongside an agentic Context-1 retrieval strategy.
• Neural Reranking: Fine-tuned cross-encoder rerankers trained with hard-negative mining sourced directly from the dense retrieval pool to maximize top-$k$ Precision and MAP.

2. Snippet Generation

• Fine-Tuned Gemma4 31b Model: Uses a Gemma model fine-tuned via LoRA in two stages:
  1. Biomedical domain adaptation for domain knowledge injection.
  2. Task-specific fine-tuning for passage extraction and retrieval-augmented generation.
• Span Extraction: Predicts and extracts relevant text spans from top-ranked documents, mapped back to precise abstract character offsets. .

3. Phase A+ / Phase B — Answer Generation

Addresses exact (Yes/No, Factoid, List) and ideal (Summary) answers using system-generated snippets (Phase A+) or gold-standard annotations (Phase B):

• Agent Quorum Mechanism: Deploys multiple LLM agents configured with diverse prompt strategies that independently propose, critique, and iteratively converge on a consensus answer.
• Adaptive Document Retention: Iteratively trims redundant or noisy context between debate rounds to keep generation focused on salient facts.
• LLM-as-a-Judge: An automated evaluation harness used during development to score candidates on factual consistency, completeness, and alignment with BioASQ task formats.

Repository Structure

• src/bioasq/phase_a/ - retrieval pipeline (BM25 + dense retrieval, reranking, query expansion).
• src/bioasq/phase_b/ - answer generation, including the agent quorum logic.
• src/bioasq/snippets/ - snippet extraction/generation.
• src/bioasq/common/ and src/bioasq/data/ - shared utilities and data handling.
• src/bioasq/cli.py - command-line entry point.
• compose.yaml, init.sql, qdrant_config.yaml - Docker Compose setup for the PostgreSQL and Qdrant services used for indexing/search.

Requirements

The project uses uv for Python dependency management (see pyproject.toml / uv.lock) and Docker Compose to run PostgreSQL (with pg_textsearch) and Qdrant.

Getting Started

1. Install dependencies with uv sync.
2. Start the supporting services: docker compose up -d.
3. Use the bioasq cli

Citation

If you use this code, please cite the associated paper:

@misc{ribeiro2026bituabioasq14bmodular,
      title={BIT.UA at BioASQ 14B: Modular Retrieval with pg_textsearch and Qdrant, and Agent-Based Answer Generation}, 
      author={André Ribeiro and Rúben Garrido and Alexander Christiansen and Richard A. A. Jonker and Sérgio Matos},
      year={2026},
      eprint={2609.04999},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.04999}, 
}