LangChain - 1.2.4 Few-shot prompt template

참조 : https://python.langchain.com/docs/modules/model_io/prompts/few_shot_examples

Few-shot prompt templates | 🦜️🔗 Langchain

참조 : https://python.langchain.com/docs/modules/model_io/prompts/few_shot_examples_chat

Few-shot examples for chat models | 🦜️🔗 Langchain

 

 

Few-shot prompt templates

 

Use Case

Few-shot prompt templates를 사용하는 경우는 아래와 같이 직접 example은 direct로 넣는 방법과 example_selector를 사용하는 방법이 있음

Step		Code	Etc.
1. Create the example set		examples = [ { 'question':'who lived longer~~~~~',      'answer':'~~~~~~'}, ~~~~ ]
2. Create a formatter		example_prompt = PromptTemplate(     input_variables=["question", "answer"],      template="Question: {question}\n{answer}" )
Using an example set	3. Feed examples and formatter	prompt = FewShotPromptTemplate(     examples=examples,     example_prompt=example_prompt,     suffix="Question: {input}",     input_variables=["input"], ) final_prompt = prompt.format(input="Who was the father of Mary Ball Washington?")
Using an  example selector	3.  Feed examples  into ExampleSelector	example_selector = SemanticSimilarityExampleSelector.from_examples(     examples,     OpenAIEmbeddings(),     Chroma,     k=1, ) question = "Who was the father of Mary Ball Washington?" selected_examples = example_selector.select_examples({"question": question})
	4.  Feed example selector	prompt = FewShotPromptTemplate(     example_selector=example_selector,     example_prompt=example_prompt,     suffix="Question: {input}",     input_variables=["input"], ) final_prompt  = prompt.format(input="Who was the father of Mary Ball Washington?")

 

 

 

Few-shot examples for chat models

Example이 몇 개 안되고 fixed 된 경우와 vectorstore에 저장된 경우를 나누어서 보면 다음과 같음

Case	Step	Code	Etc.
Fixed Examples	1.  Define examples	examples = [     {"input": "2+2", "output": "4"},     {"input": "2+3", "output": "5"}, ]
	2. Prepare prompt	example_prompt = ChatPromptTemplate.from_messages(     [         ("human", "{input}"),         ("ai", "{output}"),     ] )
	3.  few-shot prompt template	few_shot_prompt = FewShotChatMessagePromptTemplate(     example_prompt=example_prompt,     examples=examples, ) print(few_shot_prompt.format())
	4.  final prompt	final_prompt = ChatPromptTemplate.from_messages(     [         ("system", "You are a wondrous wizard of math."),         few_shot_prompt,         ("human", "{input}"),     ] )
Dynamic few-shot  prompting	1. Prepare vectorstore to select examples	examples = [     {"input": "2+2", "output": "4"},     {"input": "2+3", "output": "5"},     {"input": "2+4", "output": "6"},     {"input": "What did the cow say to the moon?", "output": "nothing at all"},     {         "input": "Write me a poem about the moon",         "output": "One for the moon, and one for me, who are we to talk about the moon?",     }, ] to_vectorize = [" ".join(example.values()) for example in examples] embeddings = OpenAIEmbeddings() vectorstore = Chroma.from_texts(to_vectorize, embeddings, metadatas=examples)
	2.  Create selector example_selector	example_selector = SemanticSimilarityExampleSelector(     vectorstore=vectorstore,     k=2, ) example_selector.select_examples({"input": "horse"})
	3. Create prompt template	few_shot_prompt = FewShotChatMessagePromptTemplate(     input_variables=["input"],     example_selector=example_selector,     example_prompt=ChatPromptTemplate.from_messages(         [("human", "{input}"), ("ai", "{output}")]     ), )
	4. final prompt	final_prompt = ChatPromptTemplate.from_messages(     [         ("system", "You are a wondrous wizard of math."),         few_shot_prompt,         ("human", "{input}"),     ] )
5. Use with an LLM		chain = final_prompt | ChatAnthropic(temperature=0.0) chain.invoke({"input": "What's 3+3?"})