yarn eval --collection <uuid> runs the hybrid retriever against a fixture of golden queries and prints precision@k / recall@k / MRR.

eval/fixtures.json is a JSON array of rows:

[
  {
    "query": "What is the SLA on response time?",
    "expectedDocumentIds": ["uuid-1", "uuid-2"]
  }
]

Each row's expectedDocumentIds is the set of documents that should appear in the top-k retrieval — any chunk from any of those documents counts as a hit.

• Precision@k: fraction of retrieved chunks whose document is in the expected set.
• Recall@k: fraction of expected documents that appear at least once in the retrieved chunks.
• MRR (Mean Reciprocal Rank): 1 / (rank of first hit), averaged across queries; 0 if no hit.

1. Upload your documents to a collection.
2. Ask the questions you'd want the system to answer; note the document IDs that should ground each answer.
3. Drop them into eval/fixtures.json (or pass --fixture path for a custom file).
4. Run yarn eval --collection <collection-uuid>.