Building Generative AI Projects

# CHAPTER 17

Building Generative AI Projects #

1. Introduction #

2. Learning Objectives #

• Architect a context-aware Customer Support Chatbot.

• Build an automated Blog Writer pipeline.

• Map out an Image Prompt Generator.

• Understand the logic for an automated Document Summarizer.

3. Beginner-Friendly Explanation #

• The Kitchen (The AI API): You don't build the kitchen; you just rent it from OpenAI. The kitchen cooks the food (generates the text).

• The Waiter (Your Python Code): This is the code you write. The waiter takes the order from the customer, adds some secret sauce (your System Prompt), walks it to the kitchen (the API call), and brings the cooked food back to the customer.

• The Menu (The Interface): This is the web page (HTML/CSS) the user interacts with.

4. Project 1: The RAG Customer Support Chatbot #

1. 1. Create a variable refundpolicy containing a 3-paragraph text detailing the strict 30-day refund rules.

1. 3. Write a robust System Prompt: *"You are a polite support bot. Read the provided refundpolicy. Answer the user's question STRICTLY using only this policy. If the user asks something unrelated, refuse to answer."*

1. 4. Create a while loop that takes user input, appends it to a message array (for memory), sends it to the API, and prints the response.

5. Project 2: The Automated SEO Blog Writer #

1. 1. Prompt 1 (The Outline): Send the topic to the API and ask it to generate a 5-point outline. Save the response.

1. 2. Prompt 2 (The Draft): Pass the Outline back to the API. Ask it to write a 200-word paragraph for each of the 5 points in a conversational tone. Save the response.

1. 3. Prompt 3 (The Formatting): Pass the Draft back to the API. Ask it to convert the draft into strict HTML format, adding <h1> tags and bullet points.

1. 4. Output the final HTML file to your computer.

6. Project 3: The Image Prompt Creator #

1. 1. User inputs: *"A dog on a skateboard"*.

1. 2. Send this to the OpenAI API with a System Prompt: *"You are an expert Midjourney prompt engineer. Take the user's simple idea and expand it into a highly descriptive 50-word prompt. Add details about cinematic lighting, the 85mm camera lens, photorealism, and the exact background environment."*

1. 3. The AI returns the complex prompt. (Optional: Pass this complex prompt directly into the DALL-E 3 API to actually generate the image!).

7. Project 4: The YouTube Video Summarizer #

1. 1. Use a free Python library (like youtube-transcript-api) to programmatically download the raw text transcript of a YouTube video using its URL.

1. 2. Pass the massive block of text to an LLM (like Claude 3 or GPT-4, which have large Context Windows).

1. 3. Set the prompt: *"Extract the 5 most actionable pieces of advice from this transcript. Format as a markdown list."*

1. 4. Output the summary.

8. Python Example: The Chained Blog Writer (Project 2) #

import openai
client = openai.OpenAI()

topic = "Benefits of Green Tea"

# Step 1: Generate Outline
outline_prompt = f"Write a 3-point outline for a blog about {topic}."
outline_response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": outline_prompt}]
).choices[0].message.content

# Step 2: Generate Full Blog based on Outline
blog_prompt = f"Here is an outline:\n{outline_response}\n\nWrite a friendly, 300-word blog post following this exact outline. Use markdown headers."
final_blog = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": blog_prompt}]
).choices[0].message.content

# Save to file
with open("green_tea_blog.md", "w") as file:
    file.write(final_blog)
print("Blog Generated and Saved!")

9. Mini Project #

10. Best Practices #

• Chaining Prompts: As demonstrated in Project 2, LLMs perform vastly better when you break a massive task (writing a blog) into smaller, sequential steps (Outline -> Draft -> Edit). This is called "Prompt Chaining."

11. Common Mistakes #

• Hardcoding Data: Don't hardcode text into your scripts if you want a portfolio app! Use input() in Python or an HTML form to allow the user to dynamically enter the topic or document they want processed.

12. Exercises #

1. 1. Explain why breaking a complex generation task into three separate API calls (Prompt Chaining) produces a higher quality output than asking the AI to do it all in one single prompt.

13. MCQs with Answers #

14. Interview Questions #

• Q: Walk me through the code architecture of a YouTube video summarizer, including how you would handle a transcript that exceeds the LLM's Context Window.

• Q: How does a developer maintain the memory of a chatbot conversation when interacting with a stateless API like OpenAI's?

15. FAQs #

16. Summary #

17. Next Chapter Recommendation #