fix: A1-14 real neural embeddings via Bumblebee multilingual-e5-small with Float32 BLOB vector cache

priv/repo/migrations/20260529115329_convert_embedding_vector_to_blob.exs

@@ -0,0 +1,33 @@
+defmodule BDS.Repo.Migrations.ConvertEmbeddingVectorToBlob do
+  use Ecto.Migration
+
+  # Embedding vectors are now persisted as a packed little-endian Float32 BLOB
+  # (VectorCacheInDb invariant) instead of JSON text. The table is a rebuildable
+  # cache and the previous lexical vectors are incompatible with the neural
+  # model, so we drop and recreate it; rows are re-embedded on next index.
+
+  def up do
+    drop table(:embedding_keys)
+    create_embedding_keys(:binary)
+  end
+
+  def down do
+    drop table(:embedding_keys)
+    create_embedding_keys(:text)
+  end
+
+  defp create_embedding_keys(vector_type) do
+    create table(:embedding_keys, primary_key: false) do
+      add :label, :integer, primary_key: true
+      add :post_id, references(:posts, column: :id, type: :string, on_delete: :delete_all),
+        null: false
+
+      add :project_id, references(:projects, type: :string, on_delete: :delete_all), null: false
+      add :content_hash, :string, null: false
+      add :vector, vector_type
+    end
+
+    create index(:embedding_keys, [:post_id])
+    create index(:embedding_keys, [:project_id])
+  end
+end