Upload 2 files

entrypoint.sh

@@ -7,6 +7,8 @@ set -euo pipefail
 export HOME=/app
 export OLLAMA_MODELS="${OLLAMA_MODELS:-/app/.ollama}"
 export OLLAMA_HOST="127.0.0.1:11434"
+CHAT_MODEL="${DEPTH_LLM_MODEL:-qwen2.5:3b}"
+EMBED_MODEL="${DEPTH_EMBED_MODEL:-nomic-embed-text}"
 
 echo "[entrypoint] starting ollama serve (models in ${OLLAMA_MODELS})..."
 ollama serve &
@@ -20,8 +22,23 @@ for i in $(seq 1 60); do
   sleep 1
 done
 
-# Log the baked models so the container log confirms the chat model is present.
-# If this list is empty, the build did not bake the model (rebuild needed).
+# Self-heal: make sure the models are present. If the image bake succeeded this
+# is instant; otherwise pull now (first cold start may take a few minutes and
+# needs outbound network). This is why the assistant works even if the build-time
+# model bake failed.
+ensure_model() {
+  if ollama list 2>/dev/null | grep -qF "$1"; then
+    echo "[entrypoint] model present: $1"
+  else
+    echo "[entrypoint] model $1 missing -> pulling now (can take minutes)..."
+    ollama pull "$1" \
+      && echo "[entrypoint] pulled: $1" \
+      || echo "[entrypoint] WARNING: could not pull $1 -- assistant will be unavailable"
+  fi
+}
+ensure_model "${CHAT_MODEL}"
+ensure_model "${EMBED_MODEL}"
+
 echo "[entrypoint] ollama models available at runtime:"
 ollama list || true