Few-shot prompt is a technique used in natural language processing (NLP) where a model is given a small number of examples (or “shots”) to learn from before generating a response or completing a task. This approach is particularly beneficial because it allows the model to understand and adapt to new tasks with minimal data, making it highly efficient and versatile. By leveraging few-shot prompts, users can achieve high-quality results without the need for extensive training datasets, thus saving time and computational resources. Additionally, this method enhances the model’s ability to generalize from limited examples, leading to more accurate and contextually relevant outputs.
One challenge of using few-shot prompts is that as the number of examples increases, the prompt can become excessively large, leading to inefficiencies and potential performance issues. To address this, a dynamic few-shot prompt technique can be employed. In this approach, a comprehensive list of prompts is stored in a vector store, and the user’s input is matched against this vector store to identify the most relevant examples. By utilizing OpenAI embeddings alongside the vector store, this method ensures that only the most pertinent examples are included in the prompt, thereby optimizing its size and relevance. This dynamic technique not only maintains the efficiency and effectiveness of few-shot learning but also enhances the model’s ability to generate accurate and contextually appropriate responses.
Architecture
The diagram above shows the overall architecture of the solution. Let's break down each component:
- Vector Store - This store will hold the few-shot prompt examples. It is indexed by each example's input and the content is the input/output pair
- Embedding Model - This model is responsible for transforming the user input into a vector, which can be used to query the vector store
- GPT Model - This is the model we will be using for chat completion, it's the one responsible for providing answers to the user
Use Case
To demonstrate how dynamic few-shot prompt can be used, let's consider a scenario where we have a chat completion that can handle 3 different types of tasks:
- Display data in a table format
- Classify texts
- Summarize texts
For each of this tasks we want to provide some examples, in order to better explain the model what it should do.
One simple way of doing this is to provide all examples related to these tasks in the prompt itself. This strategy comes with a few downsides:
- Information Overload: Too many examples can overwhelm the model, making it difficult to discern the main request.
- Confusion: The model might get confused and generate responses that are off-topic or irrelevant.
- Accuracy Issues: The focus on examples can detract from the accuracy of the response.
- Cost: More examples mean more tokens to be processed by the model, which increases cost.
It's easy to see that this strategy doesn't scale in case we need to support other types of tasks, since it would require more examples.
On the other hand, if we use dynamic few-shot prompt, we just use the most relevant examples (for the given user input) in the prompt. For an example we can decide that we only want to use the top 3 most relevant examples in the prompt.
To implement this strategy we just need the steps below:
- Define examples list
- Each example is an object that contains an input (user question example) and an output (assistant response for that question)
-
{"input": "<example_input>", "output": "<example_output>"
- Index examples list in vector store
- Index key is the embedded example's input
- Find most relevant examples
- Embed the user input using the same embedding used on the previous step
- Use the embedded user input to find the most relevant examples in the vector store
- Add examples in the prompt
- Adds example's input as a user message and its output as an assistant message
In this solution we use:
- the SemanticSimilarityExampleSelector class from langchain_core package since it already implements most of the steps mentioned above:
- Embed examples' input and index them into the vector store
- Define how many relevant examples should be returned
- Return relevant examples for a given user input
- InMemoryVectorStore
- In-memory implementation of VectorStore using a dictionary
- The package langchain-community provides other vector store implementations, like Azure Search
- AzureOpenAIEmbeddings
- Uses an openAI embedding, like text-embedding-ada-002
- The package langchain-community provides other embedding implementations, like Text2Vec
Code Implementation
The implementation of this solution is really simple. We only need 2 files: requirements.txt and main.py. Their contents can be found below:
requirements.txt
langchain-openai==0.1.21
azure-identity==1.17.1
numpy==2.0.1
main.py
# imports
from azure.identity import DefaultAzureCredential
from openai import AzureOpenAI
from langchain_openai import AzureOpenAIEmbeddings
from langchain_core.example_selectors import SemanticSimilarityExampleSelector
from langchain_core.vectorstores import InMemoryVectorStore
# constants
AZURE_OPENAI_ENDPOINT = "<AZURE_OPENAI_ENDPOINT>"
DEPLOYMENT_EMBEDDING = "<DEPLOYMENT_EMBEDDING>"
DEPLOYMENT_CHAT = "<DEPLOYMENT_CHAT>"
OPENAI_API_VERSION = "2023-05-15"
SYSTEM_MESSAGE = 'You are a helpful assistant that uses previous interactions to determine how a questions should be answered. You should try to use those same response formats'
# the examples to be indexed in the vector store. Each example should be in the format `{"input": "<example_input", "output": "<example_output>"}`
EXAMPLES_LIST = [
# data examples
{
"input": "Can you provide the population of the top 5 most populous countries?",
"output": '''
| Country | Population (2023) |
|---------------|-------------------|
| China | 1,425,000,000 |
| India | 1,417,000,000 |
| United States | 332,000,000 |
| Indonesia | 276,000,000 |
| Pakistan | 231,000,000 |
'''
},
{
"input": "What are the top 5 programming languages in 2023 by popularity?",
"output": '''
| Rank | Programming Language | Popularity (%) |
|------|----------------------|----------------|
| 1 | Python | 29.9 |
| 2 | JavaScript | 19.1 |
| 3 | Java | 16.2 |
| 4 | C# | 8.3 |
| 5 | PHP | 6.1 |
'''
},
{
"input": "Can you list the top 5 highest-grossing movies of all time?",
"output": '''
| Rank | Movie Title | Gross Revenue (USD) |
|------|-------------------------|---------------------|
| 1 | Avatar | $2.923 billion |
| 2 | Avengers: Endgame | $2.798 billion |
| 3 | Titanic | $2.195 billion |
| 4 | Star Wars: The Force Awakens | $2.068 billion |
| 5 | Avengers: Infinity War | $2.048 billion |
'''
},
{
"input": "What are the top 5 universities in the world according to the 2023 QS World University Rankings?",
"output": '''
| Rank | University | Location |
|------|-----------------------------------|----------------|
| 1 | Massachusetts Institute of Technology (MIT) | USA |
| 2 | University of Cambridge | UK |
| 3 | Stanford University | USA |
| 4 | University of Oxford | UK |
| 5 | Harvard University | USA |
'''
},
{
"input": "Can you provide the GDP of the top 5 largest economies in 2023?",
"output": '''
| Rank | Country | GDP (USD Trillions) |
|------|---------------|---------------------|
| 1 | United States | 25.3 |
| 2 | China | 17.7 |
| 3 | Japan | 4.9 |
| 4 | Germany | 4.2 |
| 5 | India | 3.5 |
'''
},
# Classification examples
{
"input": "Classify the following text: 'The quick brown fox jumps over the lazy dog.'",
"output": "Sentence"
},
{
"input": "Classify the following text: 'To be, or not to be, that is the question.'",
"output": "Quote"
},
{
"input": "Classify the following text: 'Once upon a time, in a land far, far away, there lived a young princess.'",
"output": "Story"
},
{
"input": "Classify the following text: 'E=mc^2 is a formula expressing the relationship between mass and energy.'",
"output": "Scientific Statement"
},
{
"input": "Classify the following text: 'I hope you have a great day!'",
"output": "Greeting"
},
# Summarization examples
{
"input": "Summarize the following paragraph: 'The rapid advancement of technology has significantly impacted various industries. Automation and artificial intelligence are transforming the workforce, leading to increased efficiency but also raising concerns about job displacement. Companies are investing heavily in tech to stay competitive, while governments are grappling with the need to update regulations.'",
"output": '''
- Technology is transforming industries through automation and AI.
- Increased efficiency comes with concerns about job displacement.
- Companies invest in tech; governments update regulations.
'''
},
{
"input": "Summarize the following paragraph: 'Climate change is one of the most pressing issues of our time. Rising global temperatures are causing more frequent and severe weather events, such as hurricanes and wildfires. Efforts to mitigate climate change include reducing carbon emissions, transitioning to renewable energy sources, and protecting natural habitats.'",
"output": '''
- Climate change leads to severe weather events.
- Mitigation efforts focus on reducing carbon emissions and using renewable energy.
- Protecting natural habitats is crucial.
'''
},
{
"input": "Summarize the following paragraph: 'The education system is undergoing significant reforms to better prepare students for the future. Emphasis is being placed on critical thinking, problem-solving, and digital literacy. Schools are incorporating more technology into the classroom and offering courses that align with the demands of the modern workforce.'",
"output": '''
- Education reforms focus on future readiness.
- Key skills: critical thinking, problem-solving, digital literacy.
- More technology and modern workforce-aligned courses in schools.
'''
},
{
"input": "Summarize the following paragraph: 'The healthcare industry is seeing a shift towards personalized medicine. Advances in genetic research allow for treatments tailored to individual patients, improving outcomes and reducing side effects. This approach is particularly promising in the treatment of cancer and rare genetic disorders.'",
"output": '''
- Healthcare is moving towards personalized medicine.
- Genetic research enables tailored treatments.
- Promising for cancer and rare genetic disorders.
'''
},
{
"input": "Summarize the following paragraph: 'Remote work has become increasingly common, especially after the COVID-19 pandemic. Many companies have adopted flexible work policies, allowing employees to work from home or other locations. This shift has led to changes in workplace dynamics, with a greater emphasis on digital communication and collaboration tools.'",
"output": '''
- Remote work is more common post-COVID-19.
- Companies adopt flexible work policies.
- Emphasis on digital communication and collaboration tools.
'''
}
]
# get default azure credentials. You might need to run `az login` first.
credential = DefaultAzureCredential()
# get an openai token using the default credentials. Used to authenticate with openai
openai_token = credential.get_token("https://cognitiveservices.azure.com/.default").token
# create openai client
client = AzureOpenAI(
azure_endpoint=AZURE_OPENAI_ENDPOINT,
azure_deployment=DEPLOYMENT_CHAT,
azure_ad_token=openai_token,
api_version=OPENAI_API_VERSION
)
# create openai embeddings client
embedding = AzureOpenAIEmbeddings(
azure_endpoint=AZURE_OPENAI_ENDPOINT,
deployment=DEPLOYMENT_EMBEDDING,
azure_ad_token=openai_token,
openai_api_version=OPENAI_API_VERSION
)
# creates an example selector that uses semantic similarity to retrieve examples
example_selector = SemanticSimilarityExampleSelector.from_examples(
EXAMPLES_LIST, # list of examples to index
embedding, # embedding model to use
InMemoryVectorStore, # vector store to use
k=3, # number of examples to retrieve
input_keys=["input"], # keys in the examples that contain the input text
)
while True:
# gets user input
user_input = input("Enter a sentence: ")
# select the most relevant examples based on the user input
res = example_selector.select_examples({"input": user_input})
messages = [
{"role": "system", "content": SYSTEM_MESSAGE},
]
print('retrieved examples:')
for ex in res:
# adds each example input as a user message and the corresponding output as an assistant message
messages.append({"role": "user", "content": ex['input']})
messages.append({"role": "assistant", "content": ex['output']})
print('question:{}\nanswer:{}'.format(ex['input'], ex['output']))
# adds the user input as a user message
messages.append({"role": "user", "content": user_input})
# create a completion using the messages
response = client.chat.completions.create(
messages=messages,
model="gpt-4o",
)
# print the response from the model
print('response: ', response.choices[0].message.content)
Code Execution
Pre-Requisites
- Python 3.10
- An OpenAI GPT deployment (we suggest gpt-4o model)
- An OpenAI Embedding deployment (we suggest text-embedding-ada-002 model)
- Have permission on the OpenAI service (more details here)
Install dependencies
pip install -r requirements.txt
Run
python main.py
Below you can see the results of running the code with some different types of input. Notice that on the retrieved examples section it only shows 3 examples, which are the most relevants for that input.
Enter a sentence: list me the biggest airplanes
retrieved examples:
question:Can you list the top 5 highest-grossing movies of all time?
answer:
| Rank | Movie Title | Gross Revenue (USD) |
|------|-------------------------|---------------------|
| 1 | Avatar | $2.923 billion |
| 2 | Avengers: Endgame | $2.798 billion |
| 3 | Titanic | $2.195 billion |
| 4 | Star Wars: The Force Awakens | $2.068 billion |
| 5 | Avengers: Infinity War | $2.048 billion |
question:Can you provide the GDP of the top 5 largest economies in 2023?
answer:
| Rank | Country | GDP (USD Trillions) |
|------|---------------|---------------------|
| 1 | United States | 25.3 |
| 2 | China | 17.7 |
| 3 | Japan | 4.9 |
| 4 | Germany | 4.2 |
| 5 | India | 3.5 |
question:Can you provide the population of the top 5 most populous countries?
answer:
| Country | Population (2023) |
|---------------|-------------------|
| China | 1,425,000,000 |
| India | 1,417,000,000 |
| United States | 332,000,000 |
| Indonesia | 276,000,000 |
| Pakistan | 231,000,000 |
response: Certainly! Here are some of the biggest airplanes in terms of size and capacity:
| Rank | Aircraft | Description |
|------|---------------------------------|-----------------------------------------------------------------------------------------------------------|
| 1 | Antonov An-225 Mriya | The largest cargo aircraft ever built, featuring 6 engines and a maximum takeoff weight of 640 tons |
| 2 | Airbus A380 | The largest passenger airliner, with a double-deck configuration and capacity for up to 853 passengers |
| 3 | Boeing 747-8 | One of the largest commercial aircraft, known as the "Queen of the Skies," with a capacity of 410-524 passengers |
| 4 | Lockheed C-5 Galaxy | A large military transport aircraft used by the U.S. Air Force, with a maximum takeoff weight of 381 tons |
| 5 | Boeing 777-9 | The newest variant of the Boeing 777 series, featuring a longer fuselage and higher capacity, accommodating up to 426 passengers |
These aircraft are recognized for their impressive sizes and capabilities, serving various purposes from cargo transport to long-haul passenger flights.
Enter a sentence: classify the following text: F=ma is Newton's second law of motion
retrieved examples:
question:Classify the following text: 'E=mc^2 is a formula expressing the relationship between mass and energy.'
answer:Scientific Statement
question:Classify the following text: 'The quick brown fox jumps over the lazy dog.'
answer:Sentence
question:Classify the following text: 'To be, or not to be, that is the question.'
answer:Quote
response: Scientific Statement
Enter a sentence: summarize the following text: Exercise offers numerous benefits that significantly enhance one’s quality of life. Regular physical activity helps maintain a healthy weight, reduces the risk of chronic diseases such as heart disease, diabetes, and certain cancers, and improves cardiovascular health. Additionally, exercise boosts mental health by reducing symptoms of depression and anxiety, enhancing mood, and improving sleep quality. It also increases energy levels, strengthens muscles and bones, and promotes better flexibility and balance. Overall, incorporating exercise into your daily routine can lead to a longer, healthier, and more fulfilling life. What kind of exercise do you enjoy the most?
retrieved examples:
question:Summarize the following paragraph: 'The healthcare industry is seeing a shift towards personalized medicine. Advances in genetic research allow for treatments tailored to individual patients, improving outcomes and reducing side effects. This approach is particularly promising in the treatment of cancer and rare genetic disorders.'
answer:
- Healthcare is moving towards personalized medicine.
- Genetic research enables tailored treatments.
- Promising for cancer and rare genetic disorders.
question:Summarize the following paragraph: 'The education system is undergoing significant reforms to better prepare students for the future. Emphasis is being placed on critical thinking, problem-solving, and digital literacy. Schools are incorporating more technology into the classroom and offering courses that align with the demands of the modern workforce.'
answer:
- Education reforms focus on future readiness.
- Key skills: critical thinking, problem-solving, digital literacy.
- More technology and modern workforce-aligned courses in schools.
question:Classify the following text: 'E=mc^2 is a formula expressing the relationship between mass and energy.'
answer:Scientific Statement
response: - Exercise greatly enhances quality of life.
- Benefits: healthy weight, reduced chronic disease risk, improved cardiovascular and mental health, better sleep, increased energy, stronger muscles and bones, and improved flexibility and balance.
- Leads to a longer, healthier, and more fulfilling life.
- What kind of exercise do you enjoy the most?
Next Steps
Now that you are familiar with dynamic few-shot technique you can try to take it to the next level by:
- Use a different vector store
- Use a different embedding model
- Update the examples/prompt so it can support a new task
Conclusion
The dynamic few-shot prompt technique represents a significant improvement over traditional few-shot learning. By leveraging a vector store and an embedding model, this method ensures that only the most relevant examples are included in the prompt, optimizing its relevance, size and ultimately cost. This approach not only maintains the efficiency and effectiveness of few-shot learning but also enhances the model’s ability to generate accurate and contextually appropriate responses. As a result, users can achieve high-quality outcomes with minimal data, making this technique a powerful tool for a wide range of applications.
References
Updated Sep 04, 2024
Version 1.0
franklinlindemberg
Microsoft
Microsoft
