DocumentationAPI Reference📓 Tutorials🧑‍🍳 Cookbook🤝 Integrations💜 Discord🎨 Studio (Waitlist)
Documentation

TopPSampler

Uses nucleus sampling to filter documents.

NameTopPSampler
Folder path/samplers/
Most common position in a pipelineAfter a Ranker
Mandatory input variables“documents”: A list of documents
Output variables“documents”: A list of documents

Overview

Top-P (nucleus) sampling is a method that helps identify and select a subset of documents based on their cumulative probabilities. Instead of choosing a fixed number of documents, this method focuses on a specified percentage of the highest cumulative probabilities within a list of documents. To put it simply, TopPSampler provides a way to efficiently select the most relevant documents based on their similarity to a given query.

The practical goal of the TopPSampler is to return a list of documents that, in sum, have a score larger than the top_p value. So, for example, when top_p is set to a high value, more documents will be returned, which can result in more varied outputs. The value is typically set between 0 and 1. By default, the component uses documents' score fields to look at the similarity scores.

The component’s run() method takes in a set of documents, calculates the similarity scores between the query and the documents, and then filters the documents based on the cumulative probability of these scores.

Usage

On its own

from haystack import Document
from haystack.components.samplers import TopPSampler

sampler = TopPSampler(top_p=0.99, score_field="similarity_score")
docs = [
    Document(content="Berlin", meta={"similarity_score": -10.6}),
    Document(content="Belgrade", meta={"similarity_score": -8.9}),
    Document(content="Sarajevo", meta={"similarity_score": -4.6}),
]
output = sampler.run(documents=docs)
docs = output["documents"]
print(docs)

In a pipeline

To best understand how can you use a TopPSampler and which components to pair it with, have a look at this recipe:


Related Links

See the parameters details in our API reference: