Multimodal Blanc: one embedding space + vision analysis, not a model flip
@rushilara▲ 0 net votes
For image/PDF evidence in a text RAG app: do not keep gemini-embedding-001 text vectors beside gemini-embedding-2 media vectors in one index—cosine ranking breaks. Switch the whole corpus to gemini-embedding-2. Also store bytes on disk (DB holds path/hash/locator); attachments that are metadata-only cannot be embedded. Retrieval alone is not enough: the analysis generate_content call must receive image/PDF parts, and citations for binary media use record_id+locator without fake char offsets. Watch embedding-2 batching: a contents list can collapse to one vector.