This repository was used to structure our pipeline for the BioASQ competition. Please note that some large files (e.g., models and large retrieval results) are not included here due to repository size limitations. To ensure full reproducibility, we have provided these files on Google Drive:
👉 Access all models and large retrieval files here.

Repository Structure & Project Organization

SS25AI-Project/
│
├── api_retrieval/                    # API retrieval with a pool of 1000 documents
│
├── bm_25/                            # BM25-based document ranking (filters from 1000 to 500)
│
├── NLP/                              # NLP models and utilities
│   ├── Model_In_Drive/               # Trained model storage (see Google Drive)
│   └── finetune.ipynb                # Fine-tuning code for the NLP module
│
├── Neural_Reranking/                 # Neural network based reranking
│   ├── Model_In_Drive/               # Trained model storage (see Google Drive)
│   └── Test_Batches/                 # Submission file for Phase A Test Batch 4
│
├── Phase_B/                          # Phase B: Question-specific agent modules & outputs
│   └── answers_phase_b-...json       # Example output for Phase B
│
├── .gitignore                        # Git ignore file
├── AIR_Code.ipynb                    # Main notebook. Fully developed Pipeline. 
├── SS25__BIOASQ_Phase_A_Protocol.pdf # A protocol for Phase A containing a description of the evaluation metrics as well as tutorial for official testing
├── BioASQ-task13bPhase...json        # Test dataset for the 4th batch (BioASQ challenge)
├── README.md                         # The file you are reading
├── requirements.txt                  # Python dependencies
└── training13b.json                  # Training data for model fine-tuning

Architectural Overview

This repository is organized following a two-phase pipeline:

Figure: High-level architecture of the BioASQ pipeline implemented in this repository.
Phase A handles document retrieval and ranking. Phase B processes the ranked results through question-specific agents to generate answers.

Please see the main notebook (AIR_Code.ipynb) for a pipeline overview and usage instructions.

Final Remarks

With this project our main goal was not only pure performance but rather create a pipeline which is easly replicatable and workable from others in future work. We tried to create a system which is easly understandable and tried to visualize that with the diagram above. This should make it easy for anyone to follow our work or even make it better.

Note: The API was non functional on the day of the subbmition(29.05.2025) but we hope that it will work when you try to run it. Thank you.