Bug
No validation at startup comparing stored vector dimensions against the configured embedding model dimension. If a user changes their model (e.g., 384→768), searches silently return meaningless scores.
Impact
Users switching models get garbage results with no error message. The only indication is poor search quality — not an actionable diagnostic.
Fix
- Added
getStoredDimension() to RvfBackend — reads dimension from RVF header without loading all entries
- Added validation in
ControllerRegistry.initialize() — compares stored vs configured dimension, throws EmbeddingDimensionError with actionable message in strict mode
Commit: ff50410 (ADR-0076 Track A)
ADR Reference
ADR-0076 A3
Bug
No validation at startup comparing stored vector dimensions against the configured embedding model dimension. If a user changes their model (e.g., 384→768), searches silently return meaningless scores.
Impact
Users switching models get garbage results with no error message. The only indication is poor search quality — not an actionable diagnostic.
Fix
getStoredDimension()toRvfBackend— reads dimension from RVF header without loading all entriesControllerRegistry.initialize()— compares stored vs configured dimension, throwsEmbeddingDimensionErrorwith actionable message in strict modeCommit:
ff50410(ADR-0076 Track A)ADR Reference
ADR-0076 A3