Prompts com poucos exemplos

Desenvolvimento de aplicativos de LLM com LangChain

Jonathan Bennion

AI Engineer & LangChain Contributor

Limitações dos modelos de prompt padrão

PromptTemplate + ChatPromptTemplate
✅ Lidar com poucos exemplos
❌ Não funciona com muitos exemplos
FewShotPromptTemplate

examples = [
    {
        "question": "..."
        "answer": "..."
    },
    ...
]

Criação de um conjunto de exemplo

examples = [
    {
        "question": "Does Henry Campbell have any pets?",
        "answer": "Henry Campbell has a dog called Pluto."
    },
    ...
]

# Convert pandas DataFrame to list of dicts
examples = df.to_dict(orient="records")

Formatação dos exemplos

from langchain_core.prompts import FewShotPromptTemplate, PromptTemplate


example_prompt = PromptTemplate.from_template("Question: {question}
{answer}")


prompt = example_prompt.invoke({"question": "What is the capital of Italy?"
                                "answer": "Rome"})
print(prompt.text)


Question: What is the capital of Italy?
Rome

FewShotPromptTemplate

prompt_template = FewShotPromptTemplate(

    examples=examples,

    example_prompt=example_prompt,

    suffix="Question: {input}",

    input_variables=["input"]

)

examples: a lista de dicionários
example_prompt: modelo formatado
suffix: sufixo para adicionar à entrada
input_variables

Uso do modelo de prompt com poucos exemplos

prompt = prompt_template.invoke({"input": "What is the name of Henry Campbell's dog?"})

print(prompt.text)

Question: Does Henry Campbell have any pets?
Henry Campbell has a dog called Pluto.
...

Question: What is the name of Henry Campbell's dog?

Integração com uma cadeia

llm = ChatOpenAI(model="gpt-4o-mini", api_key="...")


llm_chain = prompt_template | llm
response = llm_chain.invoke({"input": "What is the name of Henry Campbell's dog?"})

print(response.content)

The name of Henry Campbell's dog is Pluto.

Vamos praticar!

Desenvolvimento de aplicativos de LLM com LangChain