https://python.langchain.com.cn/docs/modules/model_io/prompts/prompt_templates/few_shot_examples
I. First Clarify the Core Objective of the Original Text (Use Case)
The original text defines the task at the very beginning: configure few-shot examples for “question-answering with search”. In simple terms, this means enabling the LLM to first determine “whether an intermediate follow-up question is needed” when faced with a problem, then derive the final answer step by step (e.g., when asked “Who lived longer?”, first follow up with “How old was each person when they died?” before comparing).
All code and steps are designed to achieve this task.
II. Part 1: Use the Example Set Directly (First Major Section of the Original Text)
The first step in the original text explains “using few-shot examples in prompts without LangChain templates—by manually combining examples with questions”. It is divided into 3 specific steps, with each step corresponding to the original code.
Step 1: Create a Few-Shot Example Set (Core Code from the Original Text)
The original text states: “Each example is a dictionary containing input variables”. Here, the input variables are question (the question) and answer (the answer with follow-up steps). The code is copied exactly from the original text:
# Original code: Few-shot example list (4 question-answering examples)
from langchain.prompts.few_shot import FewShotPromptTemplate
from langchain.prompts.prompt import PromptTemplate
examples = [
{
"question": "Who lived longer, Muhammad Ali or Alan Turing?",
"answer":
"""
Are follow up questions needed here: Yes.
Follow up: How old was Muhammad Ali when he died?
Intermediate answer: Muhammad Ali was 74 years old when he died.
Follow up: How old was Alan Turing when he died?
Intermediate answer: Alan Turing was 41 years old when he died.
So the final answer is: Muhammad Ali
"""
},
{
"question": "When was the founder of craigslist born?",
"answer":
"""
Are follow up questions needed here: Yes.
Follow up: Who was the founder of craigslist?
Intermediate answer: Craigslist was founded by Craig Newmark.
Follow up: When was Craig Newmark born?
Intermediate answer: Craig Newmark was born on December 6, 1952.
So the final answer is: December 6, 1952
"""
},
{
"question": "Who was the maternal grandfather of George Washington?",
"answer":
"""
Are follow up questions needed here: Yes.
Follow up: Who was the mother of George Washington?
Intermediate answer: The mother of George Washington was Mary Ball Washington.
Follow up: Who was the father of Mary Ball Washington?
Intermediate answer: The father of Mary Ball Washington was Joseph Ball.
So the final answer is: Joseph Ball
"""
},
{
"question": "Are both the directors of Jaws and Casino Royale from the same country?",
"answer":
"""
Are follow up questions needed here: Yes.
Follow up: Who is the director of Jaws?
Intermediate Answer: The director of Jaws is Steven Spielberg.
Follow up: Where is Steven Spielberg from?
Intermediate Answer: The United States.
Follow up: Who is the director of Casino Royale?
Intermediate Answer: The director of Casino Royale is Martin Campbell.
Follow up: Where is Martin Campbell from?
Intermediate Answer: New Zealand.
So the final answer is: No
"""
}
]- Key note from the original text: The
questionandanswerin each dictionary are “fixed input variable names”. The subsequent template will use these names to retrieve data, so they cannot be changed arbitrarily.
Step 2: Create a “Formatter for Individual Examples” (Called example_prompt in the Original Text)
The original text explains that this step is to “define how a single example should be displayed”—converting each dictionary (containing question + answer) in examples into text in a fixed format. The code is copied exactly from the original text:
# Original code: Formatter template for individual examples
example_prompt = PromptTemplate(
input_variables=["question", "answer"], # Must match the keys in the example dictionaries
template="Question: {question}\n{answer}" # Display format for a single example: question first, then answer
)
# Original test: Print the formatted result of the first example
print(example_prompt.format(**examples[0]))- Expected output from the original text (what you will see):
Question: Who lived longer, Muhammad Ali or Alan Turing?Are follow up questions needed here: Yes.Follow up: How old was Muhammad Ali when he died?Intermediate answer: Muhammad Ali was 74 years old when he died.Follow up: How old was Alan Turing when he died?Intermediate answer: Alan Turing was 41 years old when he died.So the final answer is: Muhammad Ali- Explanation of a difficult point:
**examples[0]is “dictionary unpacking”—it automatically passesquestionandanswerfromexamples[0]to theinput_variablesofexample_prompt. There is no need to manually writequestion=examples[0]["question"], answer=examples[0]["answer"].
Step 3: Create a Few-Shot Prompt Template (FewShotPromptTemplate)
The original text states that this step “combines the example set, formatter, and final question”. The code is copied exactly from the original text:
# Original code: Create a complete few-shot prompt template
prompt = FewShotPromptTemplate(
examples=examples, # All examples prepared in Step 1
example_prompt=example_prompt, # Formatter template for individual examples defined in Step 2
suffix="Question: {input}", # "Final question" after the examples ({input} is the question to be passed by the user)
input_variables=["input"] # Tell the template: the parameter the user needs to pass is "input" (i.e., the final question)
)
# Original test: Pass the final question and generate the complete prompt
print(prompt.format(input="Who was the father of Mary Ball Washington?"))- Expected output from the original text (what you will see):
The template will automatically format all examples according toexample_prompt, then append the question you passed:
Question: Who lived longer, Muhammad Ali or Alan Turing?Are follow up questions needed here: Yes.Follow up: How old was Muhammad Ali when he died?Intermediate answer: Muhammad Ali was 74 years old when he died.Follow up: How old was Alan Turing when he died?Intermediate answer: Alan Turing was 41 years old when he died.So the final answer is: Muhammad AliQuestion: When was the founder of craigslist born?Are follow up questions needed here: Yes.Follow up: Who was the founder of craigslist?Intermediate answer: Craigslist was founded by Craig Newmark.Follow up: When was Craig Newmark born?Intermediate answer: Craig Newmark was born on December 6, 1952.So the final answer is: December 6, 1952Question: Who was the maternal grandfather of George Washington?Are follow up questions needed here: Yes.Follow up: Who was the mother of George Washington?Intermediate answer: The mother of George Washington was Mary Ball Washington.Follow up: Who was the father of Mary Ball Washington?Intermediate answer: The father of Mary Ball Washington was Joseph Ball.So the final answer is: Joseph BallQuestion: Are both the directors of Jaws and Casino Royale from the same country?Are follow up questions needed here: Yes.Follow up: Who is the director of Jaws?Intermediate Answer: The director of Jaws is Steven Spielberg.Follow up: Where is Steven Spielberg from?Intermediate Answer: The United States.Follow up: Who is the director of Casino Royale?Intermediate Answer: The director of Casino Royale is Martin Campbell.Follow up: Where is Martin Campbell from?Intermediate Answer: New Zealand.So the final answer is: NoQuestion: Who was the father of Mary Ball Washington?- Key note from the original text:
suffixrefers to the “suffix of the examples”—i.e., the question to be asked after all examples are displayed.input_variables=["input"]means you must pass theinputparameter (your question) when callingformat.
III. Part 2: Use an Example Selector (Second Major Section of the Original Text)
The original text describes this as an “advanced usage”: instead of including all examples in the prompt, it “selects the most similar examples to the user’s question” (reducing prompt length and helping the LLM focus). It is divided into 3 steps, with each step corresponding exactly to the original code.
Step 1: Create an Example Selector (SemanticSimilarityExampleSelector)
The original text uses “semantic similarity” to select examples (e.g., if the user asks “Who was Mary Ball Washington’s father?”, it selects the example about “George Washington’s maternal grandfather” since they are most relevant). The code is copied exactly from the original text:
# Original code: Import required tools (vector store, embedding model, example selector)
from langchain.prompts.example_selector import SemanticSimilarityExampleSelector
from langchain.vectorstores import Chroma
from langchain.embeddings import OpenAIEmbeddings
# Original code: Create the example selector
example_selector = SemanticSimilarityExampleSelector.from_examples(
examples, # Still the example set from Step 1 (all selectable examples)
OpenAIEmbeddings(), # Use OpenAI’s embedding model (converts text into "numbers for similarity calculation")
Chroma, # Use Chroma vector store (stores embedded numbers for easy similarity search)
k=1 # Select only "1 most similar example"
)- Explanation of difficult points:
- “Embedding model (OpenAIEmbeddings)”: Converts text into a sequence of numbers (e.g., “apple” becomes [0.1, 0.2, …]). The more similar the numbers, the closer the semantic meaning of the text.
- “Chroma”: A tool specifically designed to store these numbers, enabling quick identification of the “example numbers” most similar to the “numbers of the user’s question”.
- “k=1”: Selects only the 1 most similar example (as used in the original text). You can also change k=2 to select 2 examples.
Step 2: Test the Example Selector (Select the Most Similar Example)
The original text uses the question “Who was the father of Mary Ball Washington?” to test which example is selected. The code is copied exactly from the original text:
# Original code: Define the test question
question = "Who was the father of Mary Ball Washington?"
# Original code: Select the most similar example based on the question
selected_examples = example_selector.select_examples({"question": question})
# Original code: Print the selected example
print(f"Examples most similar to the input: {question}")
for example in selected_examples:
print("\n")
for k, v in example.items():
print(f"{k}: {v}")- Expected output from the original text (what you will see):
The example about “George Washington’s maternal grandfather” will be selected, as both questions relate to “Mary Ball Washington’s relatives”:
Running Chroma using direct local API.
Using DuckDB in-memory for database. Data will be transient.
Examples most similar to the input: Who was the father of Mary Ball Washington?
question: Who was the maternal grandfather of George Washington?
answer:
Are follow up questions needed here: Yes.
Follow up: Who was the mother of George Washington?
Intermediate answer: The mother of George Washington was Mary Ball Washington.
Follow up: Who was the father of Mary Ball Washington?
Intermediate answer: The father of Mary Ball Washington was Joseph Ball.
So the final answer is: Joseph BallStep 3: Create a Few-Shot Template with the Example Selector
The original text states: “Replace the previous examples parameter with example_selector”—all other settings remain unchanged. The code is copied exactly from the original text:
# Original code: Create a few-shot template (using the example selector instead of all examples)
prompt = FewShotPromptTemplate(
example_selector=example_selector, # Example selector created in Step 1
example_prompt=example_prompt, # Still the formatter template for individual examples from Step 2
suffix="Question: {input}", # Still the format for the final question
input_variables=["input"] # Still the user passes the "input" parameter
)
# Original test: Pass the question and generate the complete prompt
print(prompt.format(input="Who was the father of Mary Ball Washington?"))- Expected output from the original text (what you will see):
This time, only “the 1 most similar example” is displayed (instead of all 4), with the question appended at the end:
Question: Who was the maternal grandfather of George Washington?Are follow up questions needed here: Yes.Follow up: Who was the mother of George Washington?Intermediate answer: The mother of George Washington was Mary Ball Washington.Follow up: Who was the father of Mary Ball Washington?Intermediate answer: The father of Mary Ball Washington was Joseph Ball.So the final answer is: Joseph BallQuestion: Who was the father of Mary Ball Washington?- Key note from the original text: The advantage of this approach is that “the prompt is shorter”—the LLM does not need to process irrelevant examples, leading to more accurate and faster responses.
IV. Core Summary of the Original Text (Extracted Exclusively from the Original Text)
- Two ways to use few-shot prompt templates:
- Direct use of
examples: Include all examples in the prompt (suitable for scenarios with few examples). - Use of
example_selector: Select the most relevant examples based on semantic similarity (suitable for scenarios with many examples).
- Direct use of
- 3 Essential Core Components:
- Example set (
examples): A list of dictionaries, where each dictionary contains keys corresponding toinput_variables. - Individual example formatter template (
example_prompt): Defines the display format of a single example. - Few-shot template (
FewShotPromptTemplate): Combines the example set, formatter template, and final question.
- Example set (
- Core Tools for the Example Selector:
SemanticSimilarityExampleSelector: Selects examples based on semantics.OpenAIEmbeddings: Converts text into numbers for similarity calculation.Chroma: Stores numbers and performs similarity searches.
All content is derived from the original text without adding any extra cases or code. If you have questions about any line of code, we can break it down further!