ci: pre-build FAISS indices in Docker image to fix ~60 min graph init

[luarss]

Summary

• Root cause: Graph initialisation during docker-eval CI was taking ~60 min because all 6 FAISS vector indices were rebuilt from scratch at container startup on every run. With HuggingFace CPU inference over the full corpus this is inherently slow; with Google Gemini embeddings the quota exhaustion triggers a min=60 s exponential backoff that compounds across hundreds of batches.
• Fix: Pre-build the FAISS indices inside the Docker image using HF embeddings + FAST_MODE=true. At container startup HybridRetrieverChain.create_hybrid_retriever() detects faiss_db/<name> on disk and takes the fast load_db() path, dropping graph init from ~60 min to a few seconds.
• Docker layer caching means the pre-build step is only re-run when source or dataset changes — subsequent pushes to the same PR skip it entirely.

Changes

File	What
backend/scripts/build_faiss.py	New script: calls RetrieverTools.initialize() with contextual_rerank=False (no cross-encoder download) at build time, saves all 6 indices to faiss_db/
backend/Dockerfile	Adds RUN … build_faiss.py after dataset clone; sets ENV EMBEDDINGS_TYPE=HF / ENV HF_EMBEDDINGS=thenlper/gte-large as container defaults
backend/src/agents/retriever_tools.py	Adds contextual_rerank: bool = True param to initialize() (backward-compatible default); threads it through to all 6 HybridRetrieverChain calls

How the volume interaction works

docker-compose.yml mounts faiss_data:/ORAssistant-backend/faiss_db. When the volume is empty (first run after docker compose down --volumes), Docker automatically copies the pre-built indices from the image into the volume. On subsequent starts the volume already has the indices. Either way the container finds faiss_db/<name> at startup.

Required CI-side change

Ensure backend/.env (written by ci-secret.yaml) sets:

EMBEDDINGS_TYPE=HF
HF_EMBEDDINGS=thenlper/gte-large

The runtime embedding model must match the build-time model — using GOOGLE_GENAI at runtime against HF-built indices causes a vector dimension mismatch.

Test plan

• uv run pytest tests/test_retriever_tools.py passes (24/24 ✓ locally)
• docker compose build completes and faiss_db/ is populated inside the image
• curl http://localhost:8000/conversations/ready returns {"status":"ready"} within 2–3 polling iterations (~20–30 s) after docker compose up
• docker-eval CI job finishes in < 20 min total