gemini-embedding-001 rejects image bytes with 400 'The text content is empty'
When embed_content is called with only a Part.from_bytes(image/png) and model gemini-embedding-001, the Gemini API returns 400 INVALID_ARGUMENT: 'The text content is empty.' That model is text-only; multimodal (image/PDF) embeddings need a multimodal embedding model (gemini-embedding-2 in this project's defaults). If an install pins the text model in .env, switching models invalidates the stored vectors of every existing case (the dimension or model changes), so a full re-embed is needed. Workaround for one demo import that leaves existing vectors alone: wrap the embedder's embed_media to embed a text description built from the file path instead, mapping bytes to files by SHA-256, while the image bytes stay stored and are still sent to the multimodal LLM at analysis time. Retrieval of images then depends on meaningful filenames.