[yeongseon]

Why

The existing templates/ai-agent/03_rag_metadata.py simulates vector references and LLM output. Add a separate real local retrieval-and-generation walkthrough where CUBRID remains the canonical data store.

Changes

• A localhost-only CUBRID 11.4 / Qdrant / Ollama Compose stack under templates/rag/.
• Fictional sample data ingestion, actual local embedding and vector search, parameterized CUBRID chunk retrieval, local LLM response, source IDs, and a FastAPI POST /ask.
• Deterministic offline tests included in make test, and repository index/docs updates.

Verification

• Required GitHub CI checks on final head
• Local / CI offline pytest suite
• Manual three-service end-to-end execution (not run in this session)

Scope and limitations

This is a learning example, not a security-certified public-sector AI product. It has no authentication, per-document authorization, or cross-store atomicity. Only fictional data is included; the disposable lab uses local dba defaults. Docker-based integration with real Qdrant/Ollama and CUBRID 11.2/11.4 is not claimed as verified. No existing AI-agent behavior or protected workflows were modified.

[yeongseon]

Reviewed the complete 13-file RAG demo diff at c52ef0086d3db995f196c159167c2479246b53b1 and its passing offline CI. Good separation: CUBRID stores source/chunk text, Qdrant holds vectors/IDs, Ollama runs locally; query retrieval uses parameterized CUBRID lookups and the docs correctly warn that cross-store updates are not atomic and there is no authorization. Before calling this 'runnable' and merging, please provide one actual local three-service smoke run: docker compose up, pull both models, python ingest.py, start FastAPI, POST /ask, record a real answer with matching sample:// source IDs, and check rerun ingestion. Right now all added tests use fakes and the PR explicitly marks end-to-end execution unverified; a green CI does not demonstrate CUBRID 11.4/Qdrant 1.15.5/Ollama 0.12.3 interoperability. Small reproducibility fix: the README asks a user to run python -m pytest templates/rag/tests/test_offline.py after creating a fresh venv with only pip install -r requirements.txt, but that file has only fastapi/pycubrid/uvicorn, not pytest. Document optional pip install pytest for that step (rather than making the runtime dependency heavier). Treat this as a local educational demo only, not a public-sector reference architecture or security boundary.