Retrieval / search quality benchmark

[MaximilianRau04]

Description

Evidence-grounded search is a core feature; we need a way to measure whether search actually retrieves the right passages for a given query.

Proposed Solution

Build a query/relevant-passage dataset and a benchmark measuring standard retrieval metrics (e.g. precision/recall, MRR) against it.

Acceptance Criteria

• Query set with labeled relevant passages exists
• Benchmark computes standard retrieval metrics
• Benchmark runs against a live or fixture docsgraph instance
• Results are reproducible and comparable across runs