Summary
Improve query shaping in the GBrain shadow retrieval lane before adding a heavier reranker.
Why
The first real shadow slice proved that retrieval works, but noisy literal queries can surface weaker candidates. We should normalize and shape queries before paying for a reranker.
Explore
- reduce noisy composite phrasing
- prefer concise doc-intent queries
- preserve exact title/path hints when present
- evaluate whether retrieval quality improves enough that a reranker is unnecessary
Explicit non-goal
Do not add Voyage/Voyager reranking in this issue. That stays deferred until metadata, filtering, and query shaping are all good enough to justify it.
Acceptance
- shadow retrieval uses a cleaner query form for docs search
- obvious doc-intent prompts produce more stable canonical hits
- changes remain observe-only and do not affect authoritative Cortex injection
Summary
Improve query shaping in the GBrain shadow retrieval lane before adding a heavier reranker.
Why
The first real shadow slice proved that retrieval works, but noisy literal queries can surface weaker candidates. We should normalize and shape queries before paying for a reranker.
Explore
Explicit non-goal
Do not add Voyage/Voyager reranking in this issue. That stays deferred until metadata, filtering, and query shaping are all good enough to justify it.
Acceptance