Lesson 16 — Google Gemini

Lesson 16 · Slide 1

Google Gemini

Google’s Gemini models offer unique capabilities including a 1M-token context window, native multimodal processing of images, video, audio, and documents, and tight integration with the Google ecosystem.

Learning Objectives

Understand the Gemini model family (Flash, Pro, Ultra)
Leverage the 1M-token context window for long-document tasks
Use Gemini’s multimodal capabilities (image, video, audio, documents)
Differentiate Google AI Studio from Vertex AI
Set up the google-generativeai SDK and make API calls

Lesson 16 · Slide 2

Gemini Model Family

Google offers three tiers of Gemini models, each balancing capability against speed and cost.

Model	Input (per MTok)	Output (per MTok)	Context	Best For
Gemini 2.5 Pro	$1.25–$10	$10–$30	1M	Complex reasoning
Gemini 2.5 Flash	$0.15–$0.60	$0.60–$3.50	1M	Fast, cost-effective
Gemini 2.0 Flash	Free tier available	Free tier available	1M	Prototyping

Key differentiator: Gemini 2.5 Flash offers a generous free tier (up to 500 RPD at no cost), making it excellent for prototyping and learning before committing to paid usage.

Lesson 16 · Slide 3

1M Token Context Window

Gemini’s 1M-token context window (roughly 700K words) opens use cases that were previously impossible with shorter context models.

Full codebase analysis: Load an entire repository (~50K lines) and ask questions across all files
Book-length documents: Analyse entire novels, legal contracts, or research paper collections in a single request
Long conversation history: Maintain extremely long conversations without truncation
Multi-document comparison: Compare multiple long documents side-by-side in one prompt

Cost Consideration

Filling a 1M-token context costs $1.25–$10 per request depending on the model. For repeated queries over the same document, consider chunking + RAG as an alternative. Context length is a capability, not a mandate.

Lesson 16 · Slide 4

Multimodal: Image, Video, Audio, Documents

Gemini processes multiple input modalities natively, not through separate models.

Images: Describe, analyse, OCR, compare multiple images
Video: Upload video files, get summaries, answer questions about content (unique to Gemini)
Audio: Transcribe, summarise, and answer questions about audio files
Documents: Native PDF understanding without needing to extract text first

import google.generativeai as genai

model = genai.GenerativeModel("gemini-2.5-flash")

# Analyse an image
response = model.generate_content([
    "Describe what you see in this image.",
    genai.upload_file("photo.jpg"),
])
print(response.text)

Note: Video and audio analysis are capabilities unique to Gemini among the major API providers. Claude and GPT-4o support images but have limited or no video/audio input support.

Lesson 16 · Slide 5

Google AI Studio vs Vertex AI

Google offers two platforms to access Gemini, each designed for different use cases.

Google AI Studio

Free tier available
Simple API key authentication
Best for prototyping and learning
google-generativeai SDK
Limited to Gemini models

Vertex AI

Enterprise GCP integration
IAM / service account auth
Best for production workloads
google-cloud-aiplatform SDK
Access to Claude, Llama, Mistral too

For this course, we use Google AI Studio with the google-generativeai SDK. The API surface is nearly identical; the main difference is authentication and enterprise features.

Lesson 16 · Slide 6

API Setup & SDK

Getting started with the google-generativeai SDK requires minimal setup.

pip install google-generativeai

import google.generativeai as genai
from dotenv import load_dotenv
import os

load_dotenv()
genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))

model = genai.GenerativeModel("gemini-2.5-flash")
response = model.generate_content("Explain quantum entanglement simply.")
print(response.text)
print(f"Tokens: {response.usage_metadata.prompt_token_count} in, "
      f"{response.usage_metadata.candidates_token_count} out")

Key difference from OpenAI/Anthropic: Gemini uses a GenerativeModel object rather than passing the model name to each call. You create the model once and call generate_content() on it. Chat sessions use model.start_chat().

Lesson 16 · Slide 7

Demo: Analyse a Document with Gemini

Live demonstration: use Gemini’s long context to analyse a complete document.

Demo Steps

Configure the Gemini SDK with API key
Load a long document (~50K tokens)
Send the full document as context with analysis questions
Demonstrate multi-turn conversation about the document
Show token usage metadata
Compare the same task with Claude (200K context) and note differences

Notes: Emphasise that long context is not free — show the cost of a 50K-token input request and compare with a RAG approach that might use only 2K tokens of retrieved context.

Lesson 16 · Slide 8

Lab & Homework

Lab Exercises

First Call — set up Gemini API, make a call with gemini-2.5-flash
Long Context — analyse a long text (~50K tokens)
Multimodal — describe an image using Gemini

Homework Tasks

Gemini 1M context for codebase analysis
PDF analyser with native doc understanding
Compare Gemini Flash vs Claude Haiku on 20 tasks