Dspy Knn Few Shot

Dynamic Few-Shot with dspy.KNN and dspy.KNNFewShot

Guide the user through using DSPy's KNN-based retrieval to dynamically select the most relevant few-shot demonstrations for each input at inference time, rather than using the same static examples for every query.

Step 1 — Gather context

Before generating code, ask if you do not already know:

1. How many training examples do you have? Fewer than 20 makes similarity retrieval unreliable — suggest BootstrapFewShot instead.
2. How diverse are the inputs? If all inputs are from one narrow domain, static few-shot matches KNN quality at lower overhead.
3. What embedding provider is preferred? Local sentence-transformers (free, runs locally) or an API like OpenAI embeddings (higher quality, costs money per call).
4. Do you have a quality metric? A metric lets KNNFewShot apply BootstrapFewShot filtering on top of retrieved neighbors, improving demo quality further.

What KNN and KNNFewShot are

dspy.KNN is an in-memory nearest-neighbor retriever. Given a training set and an embedding function, it converts every training example into a vector. At query time, it embeds the new input, computes dot-product similarity against all stored vectors, and returns the k most similar training examples.

dspy.KNNFewShot is an optimizer (teleprompter) that wraps KNN and BootstrapFewShot together. It compiles a student program so that every forward call first retrieves the k nearest training examples, then uses them as the few-shot demonstrations for the underlying module. The demonstrations change per input -- each query gets the examples most relevant to it.

New input ──> Embed ──> Find k nearest training examples ──> Use as demos ──> Run module

When to use

• Your training examples cover diverse subtasks and you want the LM to see only the most relevant ones for each input (e.g., a classifier that handles many categories, a QA system across different domains)
• Static few-shot examples hurt more than they help because irrelevant demos confuse the model on certain inputs
• You have enough labeled examples (at least 20-50) to make similarity-based retrieval meaningful
• You want the simplicity of few-shot prompting but with per-query adaptation

Do not use KNNFewShot when:

• You have very few training examples (< 10) -- static few-shot or BootstrapFewShot is simpler and sufficient
• All your inputs are nearly identical -- retrieval adds overhead without benefit
• You need optimized instructions, not just better demo selection -- use MIPROv2 instead

Basic usage with KNNFewShot

import dspy
from sentence_transformers import SentenceTransformer

lm = dspy.LM("openai/gpt-4o-mini")  # or any LiteLLM-supported provider
dspy.configure(lm=lm)

# 1. Prepare training data
trainset = [
    dspy.Example(question="What causes rain?", answer="Condensation of water vapor in clouds").with_inputs("question"),
    dspy.Example(question="What is photosynthesis?", answer="The process plants use to convert sunlight into energy").with_inputs("question"),
    # ... more examples
]

# 2. Set up an embedding function
encoder = SentenceTransformer("all-MiniLM-L6-v2")
embedder = dspy.Embedder(encoder.encode)

# 3. Create the optimizer
knn_optimizer = dspy.KNNFewShot(
    k=3,
    trainset=trainset,
    vectorizer=embedder,
)

# 4. Compile your module
qa = dspy.ChainOfThought("question -> answer")
optimized_qa = knn_optimizer.compile(qa)

# 5. Use it -- each call retrieves relevant demos automatically
result = optimized_qa(question="How do volcanoes form?")
print(result.answer)

Each call to optimized_qa now dynamically selects the 3 training examples most similar to the input question and includes them as few-shot demonstrations in the prompt.

Using KNN directly

If you only need the retrieval step (without the BootstrapFewShot compilation), use dspy.KNN on its own:

import dspy
from sentence_transformers import SentenceTransformer

encoder = SentenceTransformer("all-MiniLM-L6-v2")
embedder = dspy.Embedder(encoder.encode)

trainset = [
    dspy.Example(question="What is gravity?", answer="A fundamental force of attraction between masses").with_inputs("question"),
    dspy.Example(question="What is friction?", answer="A force that opposes the relative motion of surfaces").with_inputs("question"),
    # ... more examples
]

knn = dspy.KNN(
    k=3,
    trainset=trainset,
    vectorizer=embedder,
)

# Retrieve the 3 most similar examples to a new query
similar = knn(question="What is inertia?")
# similar is a list of dspy.Example objects, ranked by similarity

This is useful when you want to plug KNN retrieval into a custom module or pipeline.

Embedding configuration

KNNFewShot and KNN require a dspy.Embedder wrapping any function that takes text (or a list of texts) and returns vectors.

Using sentence-transformers (recommended default)

from sentence_transformers import SentenceTransformer

encoder = SentenceTransformer("all-MiniLM-L6-v2")
embedder = dspy.Embedder(encoder.encode)

all-MiniLM-L6-v2 is fast, small (~80MB), and works well for general-purpose similarity. For domain-specific tasks, consider models from the MTEB leaderboard.

Using OpenAI embeddings

import openai

client = openai.OpenAI()

def openai_embed(texts):
    response = client.embeddings.create(
        model="text-embedding-3-small",
        input=texts,
    )
    return [item.embedding for item in response.data]

embedder = dspy.Embedder(openai_embed)

Using any callable

Any function with the signature (str | list[str]) -> list[list[float]] works:

embedder = dspy.Embedder(my_custom_embed_function)

How retrieval works internally

1. Indexing (at init time): KNN concatenates all input fields of each training example into a single string, then calls the embedder to produce a vector per example. These vectors are stored in memory as a matrix.

2. Querying (at call time): The new input's fields are concatenated and embedded the same way. KNN computes dot-product similarity between the query vector and all stored vectors, then returns the k examples with the highest scores.

3. Demo injection (KNNFewShot only): The retrieved examples are set as the demos on each Predict module inside the compiled student program. This happens on every forward call, so demonstrations adapt per query.

The dot-product similarity means vectors should ideally be normalized (most sentence-transformer models do this by default). If your embedding function does not normalize, cosine similarity and dot-product may diverge.

Constructor parameters

dspy.KNN

dspy.KNN(
    k,           # int -- number of nearest neighbors to retrieve
    trainset,    # list[dspy.Example] -- examples to search through
    vectorizer,  # dspy.Embedder -- embedding function wrapper
)

dspy.KNNFewShot

dspy.KNNFewShot(
    k,                        # int -- number of nearest neighbors to retrieve
    trainset,                 # list[dspy.Example] -- examples to search through
    vectorizer,               # dspy.Embedder -- embedding function wrapper
    **few_shot_bootstrap_args # passed to BootstrapFewShot (e.g., metric, max_bootstrapped_demos)
)

Key method

compile(student, *, teacher=None): Returns a copy of the student program whose forward method retrieves k nearest demos per call. Accepts an optional teacher program (passed through to BootstrapFewShot).

Choosing k

Keep in mind that each retrieved demo adds to the prompt length. If your examples are long (multi-paragraph), use a smaller k to stay within context limits.

Comparison with static few-shot

Expected improvement

KNNFewShot adds the most value when inputs are diverse. On tasks with 8+ categories or domains, dynamic selection typically adds 5-15% accuracy over static few-shot (BootstrapFewShot) by ensuring each query always sees examples from relevant categories. On homogeneous inputs the gain is minimal — use BootstrapFewShot instead. Run dspy.Evaluate(devset=devset, metric=metric) on both a static and KNN-compiled program to measure the delta for your specific task before committing to the added embedding dependency.

Passing BootstrapFewShot arguments

Since KNNFewShot wraps BootstrapFewShot internally, you can pass any BootstrapFewShot parameter via **few_shot_bootstrap_args:

knn_optimizer = dspy.KNNFewShot(
    k=5,
    trainset=trainset,
    vectorizer=embedder,
    metric=my_metric,
    max_bootstrapped_demos=2,
    max_labeled_demos=3,
)
optimized = knn_optimizer.compile(my_program, teacher=teacher_program)

This retrieves 5 nearest neighbors per query and then applies BootstrapFewShot logic (with the given metric and demo limits) over those neighbors.

Gotchas

1. Claude uses unnormalized embeddings with KNN. KNN uses dot-product similarity internally, which assumes vectors are normalized. Most sentence-transformer models normalize by default, but custom embedding functions may not. If retrieval quality is poor, check whether your embedder normalizes its output. Add normalize_embeddings=True to SentenceTransformer.encode() if needed.
2. Claude forgets to call .with_inputs() on training examples. KNN concatenates only the input fields (not output fields) to create the query embedding. If .with_inputs() is missing, KNN does not know which fields are inputs and may embed the wrong thing or fail silently.
3. Claude sets k too high for long examples. Each retrieved example is injected as a few-shot demo in the prompt. With k=10 and multi-paragraph examples, the prompt can easily exceed context limits. Start with k=3 and only increase if your examples are short (single-sentence).
4. Claude uses KNNFewShot when all inputs are similar. If your inputs are homogeneous (e.g., all similar support tickets in one category), every query retrieves nearly the same demos. Static few-shot with BootstrapFewShot is simpler and equally effective in this case.
5. Claude uses OpenAI embeddings without considering cost. The embedding function runs at init time (to index all training examples) AND at every inference call. With a large trainset and high query volume, OpenAI embedding costs add up. Use local sentence-transformers (all-MiniLM-L6-v2) for cost-sensitive workloads.

Cross-references

Install any skill: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill <name>

• BootstrapFewShot for static few-shot optimization -- see /dspy-bootstrap-few-shot
• LabeledFewShot for simple static demo selection without bootstrapping -- see /dspy-labeled-few-shot
• Retrieval modules and dspy.Embedder for building RAG pipelines and embedding configuration -- see /dspy-retrieval
• Improving accuracy for the full optimization workflow -- see /ai-improving-accuracy
• For worked examples, see examples.md
• Install /ai-do if you do not have it — it routes any AI problem to the right skill and is the fastest way to work: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do

Additional resources

• dspy.KNN API docs
• dspy.KNNFewShot API docs
• DSPy optimizers overview
• For constructor signatures and method reference, see reference.md
• For worked examples (ticket classification, custom embeddings), see examples.md