Summary
The first real GBrain shadow retrieval slice works, but the next bottleneck is result intelligibility. Shadow logs need richer hit metadata so we can judge whether ranking is actually bad or just opaque.
Why
A noisy query surfaced a deprecated runbook first. That may be a ranking problem, but it is also a presentation problem. Before adding heavier reranking, we should make hits easier to interpret.
Add
- title or slug clarity
- source path
- stronger snippet/preview formatting
- timestamp metadata where available
- current/deprecated or canonical/stale signal where inferable
- keep output observe-only, no authoritative injection
Acceptance
- shadow retrieval logs include richer per-hit metadata
- enough context exists to tell canonical/current hits from stale/deprecated ones without opening the doc manually
- no changes to authoritative Cortex prompt injection path
Summary
The first real GBrain shadow retrieval slice works, but the next bottleneck is result intelligibility. Shadow logs need richer hit metadata so we can judge whether ranking is actually bad or just opaque.
Why
A noisy query surfaced a deprecated runbook first. That may be a ranking problem, but it is also a presentation problem. Before adding heavier reranking, we should make hits easier to interpret.
Add
Acceptance