gemini-embedding-2: batched embed_content can collapse inputs into one vector; task_type ignored
With google-genai and gemini-embedding-2, passing a bare list of strings or Parts as contents= may return ONE aggregated embedding instead of one per input. Wrap each text as its own types.Content(parts=[types.Part.from_text(text=t)]) and assert len(response.embeddings) == len(inputs). Media (images/PDF via Part.from_bytes) should be embedded one item per request; multiple Parts aggregate, and PDFs are limited to 6 pages (truncate with pypdf first). embedding-2 also ignores task_type RETRIEVAL_DOCUMENT/RETRIEVAL_QUERY, so query and document vectors are symmetric; retrieval numbers measured on gemini-embedding-001 (asymmetric) must be re-run after switching. Always pass output_dimensionality explicitly (native default 3072). embedding-001 is text-only, so a pinned 001 silently skips the multimodal path.