Getting Started with Gemini Pro on Google AI Studio

Gemini Pro is now broadly available, and getting started is mostly a matter of creating an API key in Google AI Studio, pasting it into a Colab notebook, and then wiring up two text models—Gemini Pro for text generation and Gemini Pro Vision for image understanding. The practical payoff is immediate: developers can test prompts in the AI Studio console, then reproduce the same behavior in code with generation settings, streaming, and safety controls.

After accepting terms in Google AI Studio, the key step is generating an API key. Users can either create a new key tied to a fresh project or create one inside an existing Google Cloud project. Once the key is copied, it’s used in Google Colab via the notebook’s secrets configuration. With the key in place, the AI Studio interface lets users try Gemini Pro and Gemini Pro Vision directly using freeform prompts, structured prompts, or chat prompts. A simple prompt—like drafting an email announcing Gemini Pro availability—returns a completion, while chat mode supports multi-turn conversation where follow-up prompts continue the context.

Beyond prompting, the console provides controls that also map cleanly into code. Temperature and other generation parameters can be adjusted, and safety settings can be edited to decide what content categories to block and how strongly to block them. Categories mentioned include harassment, hate speech, sexually explicit content, and dangerous content. The interface also provides prompt feedback, flagging whether a prompt violated any configured safety rules.

The Colab walkthrough then demonstrates the same workflow programmatically. It starts by listing available model options, including legacy bison models (text bison and chat bison), Gemini Pro, Gemini Pro Vision, and an embedding model, plus an AQA model mentioned for a future video. For text generation, the notebook uses model.generate_content with a basic question (e.g., “biggest planet”), then reads the result from response.text and prints it with Markdown. It also shows streaming by setting stream=True, which returns the output in chunks.

For more control, the notebook configures generation parameters such as temperature, top P, top K, and maximum output tokens. Safety settings are passed as a list of category dictionaries—harassment, hate speech, sexually explicit, and dangerous content—along with thresholds (for example, blocking hate speech at “medium and above”). The code then instantiates the model with both generation config and safety settings, and again checks response.prompt_feedback for objectionable content.

Chat mode is handled separately: model.startchat creates a chat session, optionally seeded with history, and chat.send_message sends new user turns while maintaining conversation state. The notebook shows how chat history is stored as role/text pairs.

Finally, Gemini Pro Vision is demonstrated using images sourced from NASA. When only an image is provided, the model returns general information about the pictured planet (Saturn). When text is added alongside the image—such as requesting the planet name and movies featuring it—the output becomes conditioned on both modalities. With two images (Earth and Saturn), the model can compare them even without explicitly naming which planet each image contains, then responds with differences and contextual facts for each.

The walkthrough closes by pointing to next steps: using Gemini Pro with LangChain and adding function calling, with those topics scheduled for follow-up videos.