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

config/config.exs

@@ -61,10 +61,15 @@ config :bds, :scripting,
   job_max_reductions: :none
 
 config :bds, :embeddings,
-  backend: BDS.Embeddings.Backends.InApp,
+  backend: BDS.Embeddings.Backends.Neural,
   model_id: "Xenova/multilingual-e5-small",
+  model_repo: "intfloat/multilingual-e5-small",
   dimensions: 384
 
+# Cache downloaded model files under the app data directory so they persist
+# across sessions (ModelCaching invariant). Overridden at runtime in prod.
+config :bumblebee, :cache_dir, Path.expand("../priv/data/models", __DIR__)
+
 config :logger, :console,
   format: "$time $metadata[$level] $message\n",
   metadata: [:request_id]

config/runtime.exs

@@ -8,4 +8,9 @@ if config_env() == :prod do
   config :bds, BDS.Repo,
     database: database_path,
     pool_size: String.to_integer(System.get_env("POOL_SIZE") || "1")
+
+  # Persist downloaded embedding model files alongside the database data dir.
+  config :bumblebee, :cache_dir,
+    System.get_env("BDS_MODEL_CACHE_DIR") ||
+      Path.join(Path.dirname(Path.expand(database_path)), "models")
 end

config/test.exs

@@ -8,3 +8,11 @@ config :bds, BDS.Repo,
   busy_timeout: 15_000
 
 config :logger, level: :warning
+
+# Tests use the deterministic lexical stub backend so the suite stays offline
+# and never downloads the ~100 MB neural model.
+config :bds, :embeddings,
+  backend: BDS.Embeddings.Backends.InApp,
+  model_id: "Xenova/multilingual-e5-small",
+  model_repo: "intfloat/multilingual-e5-small",
+  dimensions: 384