How philschmid/gemini-2.5-ai-engineering-workshop works
Project type and purpose
• This repository is an educational “AI Engineering Workshop” built around Jupyter notebooks. It is not a web app or microservice mesh, but a set of client-side notebooks that demonstrate how to use the google-genai Python SDK and Google Gemini 2.5 APIs to build agentic, multimodal AI applications.
Main components
a. User Environment
• Jupyter Notebook (local or Colab) – where the workshop is executed
• Python runtime with the google-genai SDK installed
b. Local Assets
• assets/data folder containing sample audio.mp3, book.txt, rewe_invoice.pdf, standup.mp4
• Optional: solutions notebooks (only for instructor reference)
c. SDK Client
• google-genai Python package
• Handles authentication (API key) and all RPC calls
d. Google AI Studio / Gemini Cloud Services
• Gemini 2.5 text-generation/chat endpoint
• Multimodal processing endpoints (vision, audio, video, document)
• Function-calling & tool integration endpoints
• Model Context Protocol (MCP) servers (both stdio and HTTP modes)
e. External Tools & Services (optional)
• Native code execution tool
• Google Search grounding tool
• Any user-implemented function-calling HTTP service
Relationships and data flows
• The user opens a notebook → imports google-genai → configures API key → loads local asset (e.g., audio.mp3)
• The notebook (SDK client) serializes the request, sends it over HTTPS to the appropriate Gemini service (text, vision, audio, etc.)
• Cloud service processes input, returns JSON responses (text, structured data, images, audio, transcripts)
• Notebooks parse and render outputs (charts, tables, transcripts)
• In Part 3, notebook triggers function calls: SDK forwards structured payloads to user-provided HTTP endpoints or built-in tools, then merges results back into the chat
• In Part 4, notebook launches or connects to an MCP server (stdio or HTTP) and orchestrates an always-on agentic chat loop
Architectural patterns & principles
• Client–Server: notebooks (clients) ←→ Google Gemini Cloud (servers)
• Modular demos: each notebook focuses on a capability (text, chat, multimodal, function API, MCP)
• Pluggable tools/function-calling: SDK acts as broker between model and external code/APIs
• Separation of concerns:
• Notebook = orchestration & visualization
• SDK = transport & serialization
• Cloud = model execution & heavy compute