@pytest.mark.server integration coverage that drives the ingestion
subprocess in-process and asserts the full multimodal pipeline
produces image + caption chunks with proper kind / image_path /
pair_group metadata. Default pytest runs skip; explicitly:
pytest -m server tests/python/test_rag_multimodal_integration.py
Generates a small PDF (text + PNG figure + caption paragraph) via
pymupdf, points UNSLOTH_STUDIO_HOME at tmp_path, resets the embedder
singleton, and calls _subprocess_worker with mode='multimodal'.
Asserts:
- at least one chunk of each kind (text / image / caption) is emitted
- image chunks carry a real on-disk image_path under tmp_path
- image + caption chunks share a pair_group
A second test confirms BGE-VL produces text and image vectors of the
same dimension — sanity-check for the shared-space assumption that
the multimodal retrieval path relies on.
Downloads BGE-VL-base (~600 MB) on first run, so the test is gated
behind the existing server marker rather than the default suite.
