How to Master Few-Shot Prompting with LangChain’s Example Selectors

Problem with Large Example Libraries

Hard‑coding dozens or hundreds of input‑output pairs in a prompt quickly exceeds the model's context window and raises token costs, so a method is needed to pick a small, relevant subset of examples based on the current user input.

Solution: ExampleSelector

LangChain provides the ExampleSelector component, which selects a subset of candidates according to a chosen strategy and can be combined with FewShotPromptTemplate or FewShotChatMessagePromptTemplate to build dynamic few‑shot prompts.

Common Selector Types

LengthBasedExampleSelector : selects examples sequentially until the total token length approaches a configurable max_length. Pros : simple and guarantees the prompt stays within length limits. Cons : relevance to the current input is ignored.

SemanticSimilarityExampleSelector : embeds every example with an embedding model, stores the vectors in a vector store, then computes similarity between the new input and all examples and returns the top k most similar. Pros : chooses the most relevant examples, greatly improving few‑shot effectiveness. Cons : requires an embedding model and a vector database.

MaxMarginalRelevanceExampleSelector : variant of the semantic selector that also enforces diversity among the chosen examples, preventing them from being too homogeneous. Pros : balances relevance with diverse perspectives. Cons : shares the same setup complexity as the semantic selector.

Embedding Models

Any model listed at https://docs.cherry-ai.com/knowledge-base/emb-models-info can be used.

Code Walkthrough

Example 1: LengthBased Selector (example_1_length_based_selector.py)

Steps:

Prepare five {"input": ..., "output": ...} pairs and define an example_prompt with PromptTemplate to format each pair.

Instantiate LengthBasedExampleSelector with max_length=25; the selector stops adding examples once the next one would exceed the limit.

Create a FewShotPromptTemplate that receives the selector; calling format() automatically injects the chosen examples.

examples = [
    {"input": "开心", "output": "难过"},
    {"input": "高", "output": "矮"},
    {"input": "快", "output": "慢"},
    {"input": "一个非常非常长的输入字符串，这会占用很多空间", "output": "一个同样非常非常长的输出字符串"},
    {"input": "白天", "output": "黑夜"},
]

example_selector = LengthBasedExampleSelector(
    examples=examples,
    example_prompt=example_prompt,
    max_length=25,
)

few_shot_prompt = FewShotPromptTemplate(
    example_selector=example_selector,
    example_prompt=example_prompt,
    prefix="请根据下面的示例，给出输入词的反义词。",
    suffix="输入: {user_input}
输出:",
    input_variables=["user_input"],
)

print(few_shot_prompt.format(user_input="强壮"))

Running the script prints a prompt that contains only the first three short examples, confirming that the length limit works as intended.

Example 2: Semantic Similarity Selector (example_2_semantic_similarity_selector.py)

Steps:

Read OPENAI_API_KEY from a .env file and create a HuggingFaceEmbeddings instance (e.g., model_name="mixedbread-ai/mxbai-embed-large-v1").

Use resolve_faiss() to import the FAISS vector store; if the dependency is missing, the script prints an installation hint.

Build a SemanticSimilarityExampleSelector via from_examples, passing the examples, the embeddings, the FAISS class, and k=2 to retrieve the two most similar examples.

Pass this selector to FewShotPromptTemplate; when format() is called, the nearest examples are automatically injected into the prompt.

embeddings = HuggingFaceEmbeddings(model_name="mixedbread-ai/mxbai-embed-large-v1")
example_selector = SemanticSimilarityExampleSelector.from_examples(
    examples,
    embeddings,
    faiss_cls,
    k=2,
)

few_shot_prompt = FewShotPromptTemplate(
    example_selector=example_selector,
    example_prompt=example_prompt,
    prefix="请根据下面的示例，为输入生成一个富有想象力的场景描述。",
    suffix="输入: {user_input}
输出:",
    input_variables=["user_input"],
)

final_prompt = few_shot_prompt.format(user_input="一头孤独的狼")
print(final_prompt)

If faiss-cpu, tiktoken, or the API key are missing, the script prints a clear message indicating which dependency must be installed or configured.

Dependencies

faiss-cpu

: Facebook AI’s high‑performance similarity search library, used via langchain_community.vectorstores.FAISS. Runs on CPU, no GPU required. tiktoken: OpenAI’s fast tokenizer; LangChain relies on it for accurate token‑length calculations when building or trimming prompts.

References

How to: use example selectors – https://python.langchain.com/docs/how_to/example_selectors

How to: select examples by length – https://python.langchain.com/docs/how_to/length_based_example_selector

How to: select examples by semantic similarity – https://python.langchain.com/docs/how_to/semantic_similarity_example_selector

How to: select examples by semantic n‑gram overlap – https://python.langchain.com/docs/how_to/semantic_ngram_overlap_example_selector

How to: select examples by maximal marginal relevance – https://python.langchain.com/docs/how_to/maximal_marginal_relevance_example_selector

How to: select examples from LangSmith few‑shot datasets – https://python.langchain.com/docs/how_to/langsmith_few_shot_datasets