🧠 Few-Shot vs Zero-Shot Prompting in Generative AI (with OpenAI Python SDK)

📘 Introduction

In the world of prompt engineering, how you design your input to an LLM (like OpenAI’s GPT models) greatly affects the output. Two foundational techniques are:

• Zero-shot prompting: Ask the model to perform a task without any examples.

• Few-shot prompting: Provide the model with a few examples of the task, so it can mimic the pattern.

These techniques are essential when you're using LLMs for text classification, generation, summarization, translation, or any custom NLP task.

🎯 Zero-Shot Prompting

In zero-shot prompting, you give the model only the instruction, and it relies on its pre-trained knowledge to generate a response.

✅ Example Task: Sentiment Classification

from openai import OpenAI
client = OpenAI()

prompt = "Classify the sentiment of the following sentence as Positive, Negative, or Neutral:\n\n'Stunning visuals but the plot was boring.'"

response = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[{"role": "user", "content": prompt}],
    temperature=0.0,
)

print(response.choices[0].message.content)

🧾 Output:

Negative

🔍 Why it works: The model was trained on a huge corpus of sentiment-labeled examples and can apply that knowledge even if we don’t provide training examples in the prompt.

🧪 Few-Shot Prompting

Here, we include a few labeled examples in the prompt so the model understands the format, structure, and expectations better.

✅ Example: Same Sentiment Classification, Now Few-Shot

few_shot_prompt = """
Classify the sentiment of each sentence as Positive, Negative, or Neutral.

Example 1:
Sentence: "I love this movie, it was fantastic!"
Sentiment: Positive

Example 2:
Sentence: "The product is okay, nothing special."
Sentiment: Neutral

Example 3:
Sentence: "Worst customer service I've ever experienced."
Sentiment: Negative

Now classify this sentence:
Sentence: "Stunning visuals but the plot was boring."
Sentiment:
"""

response = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[{"role": "user", "content": few_shot_prompt}],
    temperature=0.0,
)

print(response.choices[0].message.content)

🧾 Output:

Negative

🧠 Why it works better sometimes: The model picks up the input-output format and the task intent more clearly, improving accuracy—especially for domain-specific tasks.

🔍 When to Use Which?

Technique Best For Zero-shot Simple or general tasks where model knowledge is sufficient Few-shot Custom formats, domain-specific tasks, or when extra guidance is needed

💡 Tip: If performance is still lacking, consider fine-tuning or using tool-augmented prompting (like RAG, LangChain).

🧰 OpenAI Python SDK Setup

To use the examples above, ensure you’ve set up OpenAI:

pip install openai

And set your API key:

import os
os.environ["OPENAI_API_KEY"] = "your_api_key_here"

💡 Bonus: Use f-strings for dynamic few-shot examples

examples = [
    ("I love this!", "Positive"),
    ("Not worth the money.", "Negative"),
    ("It's just okay.", "Neutral")
]

new_sentence = "The acting was good but the story lacked depth."

few_shot_dynamic = "Classify sentiment as Positive, Negative, or Neutral.\n\n"
for i, (s, label) in enumerate(examples, 1):
    few_shot_dynamic += f"Example {i}:\nSentence: \"{s}\"\nSentiment: {label}\n\n"

few_shot_dynamic += f"Now classify this sentence:\nSentence: \"{new_sentence}\"\nSentiment:"

response = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[{"role": "user", "content": few_shot_dynamic}],
    temperature=0.0,
)

print(response.choices[0].message.content)

📌 Final Thoughts

• Zero-shot is fast, simple, and surprisingly effective for common tasks.

• Few-shot gives you more control, especially when task format matters.

• These are the foundations of prompt engineering—mastering them sets you up for advanced tools like LangChain, LangGraph, or Retrieval-Augmented Generation (RAG).