This repository contains the evaluation queries used in the paper "Bayesium: An Agentic RAG Interface for Traceable Exploration of Bayesian Network Literature" (PGM 2026).
The evaluation is split into four components:
- Routing (100 queries): 25 BN, 24 MD, 26 CC, 25 OOC
- Retrieval (384 gold chunks): Stratified sample of 384 paragraphs (96 per document-length quartile) with synthetically generated questions
- Generation (50 BN queries): Conceptual questions about Bayesian networks
- Text-to-SQL (18 queries): Bibliographic metadata queries
routing_queries.json # 100 hand-written routing queries with gold labels
retrieval_queries.json # 384 (chunk_id, synthetic_question) pairs
generation_queries.json # 50 BN conceptual questions
text_to_sql_queries.json # 18 metadata filter queries
Each JSON file follows the format:
[
{
"id": 1,
"query": "What is d-separation in Bayesian networks?",
"gold_category": "BN"
},
...
][
{
"id": 1,
"chunk_id": 12345,
"synthetic_question": "How does the K2 algorithm handle structure learning?",
"doc_id": 789,
"doc_title": "Learning Bayesian Network Structures"
},
...
][
{
"id": 1,
"query": "What is the K2 score and how is it used to learn Bayesian network structure?"
},
...
][
{
"id": 1,
"query": "papers by Pearl on causality after 1995",
"gold_sql": "SELECT doc.id, doc.title, doc.year FROM doc JOIN doc_author ON doc.id = doc_author.doc_id JOIN author ON doc_author.author_id = author.id WHERE author.last_name = 'Pearl' AND doc.year > 1995 AND doc.title ILIKE '%causal%'",
"gold_result_count": 12
},
...
]If you use these queries, please cite:
@inproceedings{paniagua2026bayesium,
title = {Bayesium: An Agentic RAG Interface for Traceable Exploration of Bayesian Network Literature},
author = {Paniagua, Kevin and Raposo, Adrián and Larrañaga, Pedro and Bielza, Concha},
booktitle = {Proceedings of Machine Learning Research (Probabilistic Graphical Models)},
year = {2026}
}These queries are released under CC BY 4.0.