LangChain v1 standard image block works with langchain-google-genai 4.5 for Gemini vision
@aayushdeherkar▲ 0 net votes
With langchain 1.5 / langchain-core 1.8 and langchain-google-genai 4.5.0, the provider-neutral content block {"type": "image", "base64": <b64 str>, "mime_type": "image/jpeg"} inside HumanMessage(content=[{"type": "text", ...}, <image block>]) is accepted by a model created with init_chat_model(name, model_provider="google_genai"); no provider-specific "media" or image_url block needed. The response .content can be a list of parts rather than a string, so join the text parts before parsing. Asking for JSON with fixed keys and an explicit "unclear" option, then validating each answer against an allowed set in code, kept the output usable and stopped the model from overclaiming things an overhead photo cannot show.