# The winner of all_solutions_multihop: 'hybrid_rerank_default'. Run it with `ragbench run --config <this file>`.
# `dataset:` is the benchmark's data; replace it with your production documents before deploying.
run:
  name: all_solutions_multihop_winner
  output_dir: results
dataset:
  documents_path: data/multihop-rag/docs
  questions_path: data/multihop-rag/questions.jsonl
  qrels_path: data/multihop-rag/qrels.jsonl
systems:
- type: hybrid_rerank
  name: hybrid_rerank_default
  chunker:
    type: word
    chunk_size: 500
    chunk_overlap: 80
  retrieval:
    vector_store: numpy
    bm25_top_k: 20
    vector_top_k: 20
    candidate_top_k: 30
    final_top_k: 5
    rrf_k: 60
    reranker: local_relevance
    multi_query: true
    max_query_variants: 4
  models:
    embedding: text-embedding-3-small
    generator: gpt-6-luna
openrouter:
  provider:
    data_collection: deny
limits:
  tokens_per_minute: 900000
evaluation:
  k_values:
  - 1
  - 3
  - 5
  - 10
  judge_enabled: true
  judge:
    panel:
    - anthropic:claude-haiku-5-5
    - gpt-6-luna
    independent: true
    scorers: first
  trials: 1
  max_questions: null
  max_workers: 4
  max_cost_usd: 24.0