Multi-Step Model Workflows

Complex Logic > Single Prompt.

The Waterfall Pattern

Imagine generating a Newsletter.

1. Fetcher: Scrapes 5 URLs.
2. Summarizer: Runs Gemini Flash on each URL. (List[Returns]).
3. Synthesizer: Runs Gemini Pro on the combined summaries to write the intro.
4. Editor: Runs Gemini Pro to check for grammar.

Managing State

You need a way to pass data.

• Simple: Passing variables in a Python script.
• Robust: Using a framework like LangGraph or Airflow.

Parallelism

Step 2 (Summarize 5 URLs) can happen in Parallel.

• Don't loop for url in urls: call_gemini(). That takes 5x time.
• Use asyncio.gather() to call all 5 at once.

import asyncio

async def summarize(text):
    return await model.generate_content_async(text)

async def main():
    tasks = [summarize(url_text) for url_text in urls]
    summaries = await asyncio.gather(*tasks)

Summary

Think of AI operations as an assembly line. Specialized workers (models) doing specific steps, often in parallel.

In the next lesson, we handle Data Processing.