AICORE

Embedding Model Switch

An operator changes the embedding model from a Settings dropdown, is told what it costs, and a guided procedure rebuilds vectors and indexes and activates the model only when verified, with vector search off meanwhile and a guard that disables it on any mismatch.

State
Done
Progress
100%
Steps
65 / 65
Started
2026-10-05
Finished
2026-10-05
Modules
LaraAI, LaraCore

Tasks.

  • Task 1: Dimensions come from the profile, and the active profile can be overridden for a call5/5
  • Task 2: The dimension probe and ai:embeddings:probe5/5
  • Task 3: Stored vectors are checked, and the embedding job takes a profile5/5
  • Task 4: Command-managed settings (Core)5/5
  • Task 5: The vector availability guard (Core)5/5
  • Task 6: The embedding settings, the active profile from a setting, and the AI side of the guard5/5
  • Task 7: The setting-change confirmation hook (Core)5/5
  • Task 8: The alert, the switch start and its preflight5/5
  • Task 9: Embeddings, indexes, verification and activation5/5
  • Task 10: PostgreSQL with pgvector5/5
  • Task 11: Documentation, spec amendments and closing5/5
  • Task 12: Index documents carry the vectors of one model (Core)5/5
  • Task 13: The serving model is Core's search.vector.model only (AI)5/5

Read the full plan on GitHub