Few-Shot Prompting: Learning in Real-Time

Sometimes, no matter how much "System Instruction" you give a model, it still fails to follow your specific style. In these cases, you use Few-Shot Prompting. You provide the model with 3 to 5 examples of "Input -> Output" pairs so it can hallucinate the pattern.

1. The Anatomy of a Few-Shot Prompt

from langchain.prompts import FewShotPromptTemplate, PromptTemplate

# 1. Define the format for ONE example
example_prompt = PromptTemplate.from_template("Input: {input}\nOutput: {output}")

# 2. Provide the EXAMPLES
examples = [
    {"input": "happy", "output": "sad"},
    {"input": "tall", "output": "short"}
]

# 3. Assemble the FewShot Template
prompt = FewShotPromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
    suffix="Input: {word}\nOutput:",
    input_variables=["word"]
)

print(prompt.format(word="hot"))
# Output: "Input: happy\nOutput: sad\nInput: tall\nOutput: short\nInput: hot\nOutput:"

---

2. When to use Few-Shotting

• Sentiment Analysis: "Is this comment positive? Example A: Yes..."
• Code Generation: Show the model exactly how you want your variable names formatted.
• Tone Consistency: If your agent must speak like a specific brand or character.

---

3. Dynamic Few-Shotting (Advanced)

What if you have 1,000 examples? You can't put all of them in the prompt; it would be too expensive and the model would get confused.

• We use an Example Selector (often powered by a Vector DB - Module 15) to find the 2-3 most "relevant" examples for the current user query.

---

4. Visualizing the Pattern Match

graph LR
    User[Query: 'Fast'] --> Selection[Pick 2 antonym examples]
    Selection --> Final[Big Prompt with examples]
    Final --> LLM[Pattern Recognition]
    LLM --> Result[Output: 'Slow']

---

5. Engineering Tip: Quality over Quantity

It is better to have 3 perfect examples than 20 mediocre examples. If your examples are sloppy, the agent's output will be sloppy.

---

Key Takeaways

• Few-Shotting is the most powerful way to control output format and style.
• PromptTemplate is used to define how each individual example is printed.
• The Suffix is where the actual user query is placed.
• Use Dynamic Selection if your example database is large.