build: enable multimodal serving on cloud run

Bake the ColQwen2 weights so startup pays no HF fetch. The 2B CPU
query encoder (~8GB fp32) plus bge-m3 needs 16Gi/4vCPU; 8Gi was too
small.

Author: NorthernLightx

.github/workflows/deploy.yml

@@ -81,7 +81,11 @@ jobs:
           # config (corpus collection, pages dir, etc.) is preserved. Enables the
           # experimental DCI mode; with no server key it runs on the caller's key.
           # max-instances=1 + timeout bound the worst-case Cloud Run burn rate.
-          flags: --port=8000 --memory=8Gi --cpu=2 --cpu-boost --min-instances=0 --max-instances=1 --timeout=120 --allow-unauthenticated --update-env-vars=RAG_ENABLE_DCI=true
+          # ADR 0028: 16Gi/4vCPU fits the in-process ColQwen2 query encoder (2B,
+          # fp32 ~8GB) alongside bge-m3; Cloud Run requires >=4 vCPU at 16Gi.
+          # RAG_ENABLE_MULTIMODAL turns the visual leg on (it stays inert unless
+          # the visual collection is populated in the baked Qdrant snapshot).
+          flags: --port=8000 --memory=16Gi --cpu=4 --cpu-boost --min-instances=0 --max-instances=1 --timeout=120 --allow-unauthenticated --update-env-vars=RAG_ENABLE_DCI=true,RAG_ENABLE_MULTIMODAL=true
 
       - name: Print URL
         run: |

Dockerfile

@@ -70,6 +70,16 @@ COPY --chown=app:app qdrant_local /home/app/qdrant_local
 RUN /home/app/.venv/bin/python -c \
     "from sentence_transformers import SentenceTransformer, CrossEncoder; SentenceTransformer('BAAI/bge-m3'); CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')"
 
+# ADR 0028: bake the ColQwen2 visual encoder so the multimodal serve path pays
+# no HuggingFace fetch at startup — a ~4 GB cold download would blow the Cloud
+# Run startup window. from_pretrained caches the adapter, its Qwen2-VL-2B base,
+# and the processor. Only the query is encoded at serve time; the page vectors
+# are pre-built into the Qdrant snapshot (scripts/build_visual_index.py). Adds
+# ~4 GB to the image; the visual leg only activates when RAG_ENABLE_MULTIMODAL
+# is set AND the visual collection is populated.
+RUN /home/app/.venv/bin/python -c \
+    "from colpali_engine.models import ColQwen2, ColQwen2Processor; ColQwen2.from_pretrained('vidore/colqwen2-v1.0'); ColQwen2Processor.from_pretrained('vidore/colqwen2-v1.0')"
+
 # RAG_RERANKER_MODEL: light CPU-feasible cross-encoder (see settings.py).
 # RAG_RERANK_TOP_K: trim the rerank pool 50 -> 20 for CPU latency; the 20-paper
 # demo corpus doesn't need a 50-candidate pool.