Few-Shot Prompting

The limits of describing what you want

Every template so far has told the AI what to do in words: "answer in two sentences," "you are a pirate." That works for behavior, but it's clumsy for describing an exact format, especially a format that's easier to show than to describe. Try writing instructions in plain English for "summarize a sentence as a single all-caps word matching the main action", and you'll notice it's fiddly to get exactly right. It's much easier to just show a few examples and let the pattern speak for itself.

That's few-shot prompting: instead of (or in addition to) describing the task, you show the model a handful of correctly-done examples, then give it a new, real input and let it continue the pattern.

One example, formatted

example_prompt = ChatPromptTemplate.from_messages(
    [
        ("human", "{input}"),
        ("ai", "{output}"),
    ]
)

This is a small template (same idea as every template since Lesson 3) that turns one {"input": ..., "output": ...} pair into a fake human/ai exchange, a pretend little snippet of conversation where the human asked something and the ai answered exactly the way we want.

Many examples, expanded automatically

examples = [
    {"input": "The cat is sleeping on the warm windowsill.", "output": "SLEEPING"},
    {"input": "She sprinted across the finish line first.", "output": "RUNNING"},
    {"input": "He is reading a mystery novel by the fire.", "output": "READING"},
]

few_shot_prompt = FewShotChatMessagePromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
)

FewShotChatMessagePromptTemplate runs every dictionary in examples through example_prompt, producing one fake human/ai exchange per example, all strung together in order. Three examples in, three human/ai pairs come out.

Assembling the final template

final_prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "Summarize the main action in the sentence with one word."),
        few_shot_prompt,
        ("human", "{input}"),
    ]
)

Notice few_shot_prompt is dropped directly into the messages list, right alongside plain ("system", ...) and ("human", ...) tuples. ChatPromptTemplate is fine mixing a block of expanded examples in with regular messages. The very last entry, ("human", "{input}"), is the real question, still just a blank, waiting to be filled with something new.

What the model actually receives

Run the lesson and look at the printed conversation:

[system] Summarize the main action in the sentence with one word.
[human] The cat is sleeping on the warm windowsill.
[ai] SLEEPING
[human] She sprinted across the finish line first.
[ai] RUNNING
[human] He is reading a mystery novel by the fire.
[ai] READING
[human] The children were laughing loudly in the playground.

The model reads this the same way it read the memory example back in Lesson 2: as one long conversation. It sees three rounds of "human asks, ai answers in a specific style," then a fourth human message with no answer yet. The most natural continuation, given everything it just read, is to answer in that same style: LAUGHING. Nobody told it "respond in all caps with one word" in plain English, it inferred the format purely from the pattern.