Module 10 Lesson 3: Creating a Simple Workflow

In the last two modules, we learned about individual skills: RAG for reading facts, and Tools for taking actions. But in the real world, you don't just use one. You use both in a specific sequence. This is called Orchestration.

In this lesson, we will build a conceptual "Travel Agent" workflow to see how these parts work together in a loop.

---

1. The Design of the Workflow

Imagine a user says: "I want to go to Paris. Do I have enough money in my budget for a 5-day trip?"

To answer this, the AI needs to:

1. Retrieve Data (RAG): Read the user's private holiday budget document.
2. Call a Tool: Call a flight API to get the current price of a ticket to Paris.
3. Perform Logic: Subtract the flight cost from the budget and calculate if the remainder covers 5 days of hotels.
4. Respond: Give the final "Yes" or "No" to the user.

---

2. Managing "State" and Memory

The biggest challenge in complex workflows is State. The model needs to remember what happened in Step 1 while it is doing Step 3.

• In professional apps, we don't just rely on the model's built-in context window. We often save these values in a Database and feed them back to the model as "Memory" for each new step.

---

3. The Multi-Step Loop

graph TD
    User["User Query"] --> Step1["Step 1: RAG Search (User Data)"]
    Step1 --> Step2["Step 2: Tool Call (Flight Prices)"]
    Step2 --> Step3["Step 3: Reasoning (Math/Logic)"]
    Step3 --> Step4["Step 4: Safety Check (Guardrails)"]
    Step4 --> Final["Final Output to User"]

---

4. Why Frameworks (like LangChain or AutoGen) exist

Building these loops manually is difficult. Developers use Orchestration Frameworks to handle the "glue" between the steps. These frameworks:

• Automatically format the JSON for tool calls.
• Handle errors (e.g., if a tool fails, the AI is told to try again).
• Manage the conversation history so the context doesn't get cluttered.

---

Lesson Exercise

Goal: Build a logic flow.

You are building an AI "Tutor" for a math student.

1. Step 1: The student asks a question.
2. Step 2: The AI uses RAG to read the textbook to find the relevant formula.
3. Step 3: The AI uses a Tool (Calculator) to verify the numbers.
4. Step 4: The AI uses Prompt Engineering to explain the answer simply.

Challenge: What happens if the RAG search in Step 2 fails to find the formula? Write a "Fallover" instruction for the AI. (e.g., "If you can't find it, ask the student for its name.")

---

Summary

In this lesson, we established:

• Professional apps link multiple LLM capabilities (RAG, Tools, Memory) into workflows.
• Orchestration is the "glue" that ensures steps happen in the correct order.
• Frameworks help manage the complexity of multi-step AI agents.

Next Module: We face the reality check. In Module 11: Limitations of LLMs, we'll learn about what AI can't do, from long-horizon planning to understanding causality.