fix: CPU-only torch + move indexing to container startup

- Install torch from CPU wheel index to prevent sentence-transformers
from pulling multi-GB CUDA wheels (nvidia-cublas, cudnn, etc.) onto
the CPU-only HF Space, which was OOM-killing the build.
- Move repo indexing out of docker build (which loaded the embedding
model + embedded two repos into RAM) into a startup entrypoint.
First container start now indexes on demand; subsequent restarts
reuse ephemeral disk.
- Multi-stage build: node:20-slim for frontend, python:3.11-slim for
app. Frontend dist copied across.

Fixes HF Space build error (exit 137 / OOMKilled).

Dockerfile

@@ -6,7 +6,7 @@ RUN npm ci
 COPY frontend/ ./
 RUN npm run build
 
-# ---- Stage 2: Python app + pre-built indexes -------------------------------
+# ---- Stage 2: Python app ----------------------------------------------------
 FROM python:3.11-slim
 
 # git is needed for `git diff`-based incremental indexing and to clone the
@@ -18,6 +18,13 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 
 WORKDIR /app
 
+# CPU-only torch first — sentence-transformers will see it already installed
+# and skip the multi-GB CUDA wheels (this Space has no GPU; the CUDA download
+# was OOM-killing the build).
+RUN pip install --no-cache-dir \
+    --index-url https://download.pytorch.org/whl/cpu \
+    torch==2.12.1
+
 COPY requirements.txt ./
 RUN pip install --no-cache-dir -r requirements.txt
 
@@ -27,22 +34,25 @@ COPY . .
 COPY --from=frontend /frontend/dist ./frontend/dist
 
 # HF Spaces containers run as a non-root user by convention; create one and
-# make sure it owns everything, including the HF model cache written below.
+# make sure it owns everything, including the HF model cache and the index
+# directory written at container startup.
 RUN useradd -m -u 1000 appuser
 ENV HF_HOME=/app/.cache/huggingface
-RUN mkdir -p "$HF_HOME" && chown -R appuser:appuser /app
+RUN mkdir -p "$HF_HOME" /app/indexes /app/showcase_repos && \
+    chown -R appuser:appuser /app
 
 USER appuser
 ENV PYTHONPATH=/app/src
 ENV REPOGRAG_INDEX_DIR=/app/indexes
 
 # Showcase repo used for the second demo entry in indexes/showcase.json.
-RUN git clone --depth 1 https://github.com/expressjs/express.git /app/showcase_repos/express
+RUN git clone --depth 1 https://github.com/expressjs/express.git \
+    /app/showcase_repos/express
 
-# Build both indexes at image-build time (bakes in the ~320MB embedding model
-# download too) so the first real user request isn't slow.
-RUN python3 -m repograg.cli index . --db /app/indexes/repograg.db && \
-    python3 -m repograg.cli index /app/showcase_repos/express --db /app/indexes/express.db
+# Indexing (embedding model download + embedding two repos) happens at
+# container startup, not at build time — the builder VM was OOM-killing it.
+COPY --chown=appuser:appuser docker-entrypoint.sh /app/docker-entrypoint.sh
+RUN chmod +x /app/docker-entrypoint.sh
+CMD ["/app/docker-entrypoint.sh"]
 
 EXPOSE 7860
-CMD ["uvicorn", "repograg.server:app", "--host", "0.0.0.0", "--port", "7860"]

docker-entrypoint.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+set -e
+if [ ! -f /app/indexes/repograg.db ]; then
+    echo "Building repograg index..."
+    python3 -m repograg.cli index . --db /app/indexes/repograg.db
+fi
+if [ ! -f /app/indexes/express.db ]; then
+    echo "Building express index..."
+    python3 -m repograg.cli index /app/showcase_repos/express \
+        --db /app/indexes/express.db
+fi
+exec uvicorn repograg.server:app --host 0.0.0.0 --port 7860