fix: pre-download GGUF and embedding models at build time to eliminate cold-start delay

Dockerfile

@@ -8,26 +8,28 @@ RUN apt-get update && apt-get install -y \
 
 WORKDIR /app
 
-# Set at build time so pip uses them when compiling llama-cpp-python
 ENV CMAKE_ARGS="-DGGML_BLAS=OFF -DGGML_METAL=OFF -DGGML_OPENMP=OFF -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=OFF"
 ENV CMAKE_BUILD_PARALLEL_LEVEL=4
+ENV HF_HUB_CACHE=/app/model
 
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
 COPY . .
 
-# Pre-download the GGUF model at build time so cold starts don't fetch ~2.5 GB at runtime
-RUN python - <<'EOF'
-from huggingface_hub import hf_hub_download
-hf_hub_download(
-    repo_id="Jackrong/Qwen3.5-4B-Neo-GGUF",
-    filename="Qwen3.5-4B.Q4_K_S.gguf",
-    cache_dir="./model",
-)
-EOF
-
-RUN useradd -m -u 1000 user
+# Pre-download GGUF model (~2.5 GB) at build time
+RUN python -c "\
+from huggingface_hub import hf_hub_download; \
+hf_hub_download(repo_id='Jackrong/Qwen3.5-4B-Neo-GGUF', filename='Qwen3.5-4B.Q4_K_S.gguf', cache_dir='/app/model')"
+
+# Pre-download embedding model at build time
+RUN python -c "\
+from huggingface_hub import snapshot_download; \
+snapshot_download('BAAI/bge-small-en-v1.5')"
+
+RUN useradd -m -u 1000 user \
+    && chown -R 1000:1000 /app/model
+
 USER user
 ENV HOME=/home/user PATH=/home/user/.local/bin:$PATH
 

src/utils.py

@@ -17,12 +17,17 @@ class UTF8LocalFileSystem(LocalFileSystem):
 
 def resolve_gguf_model_path(status_callback=None) -> str:
     """Return local path to the GGUF model file, downloading from HF Hub if needed."""
-    direct = Path("model") / MODEL_FILENAME
+    import os
+
+    # Respect HF_HUB_CACHE env var (set in Dockerfile to /app/model)
+    cache_dir = Path(os.environ.get("HF_HUB_CACHE", "./model"))
+
+    direct = cache_dir / MODEL_FILENAME
     if direct.exists():
         return str(direct)
 
     snapshots_root = (
-        Path("model")
+        cache_dir
         / f"models--{'--'.join(MODEL_REPO_ID.split('/'))}"
         / "snapshots"
     )
@@ -35,5 +40,5 @@ def resolve_gguf_model_path(status_callback=None) -> str:
     return hf_hub_download(
         repo_id=MODEL_REPO_ID,
         filename=MODEL_FILENAME,
-        cache_dir="./model",
+        cache_dir=str(cache_dir),
     )