Principle:Marker Inc Korea AutoRAG Passage Sampling

Knowledge Sources	AutoRAG Docs
Domains	NLP, Data_Sampling
Last Updated	2026-02-08 06:00 GMT

Overview

A sampling strategy that selects corpus passages as ground-truth retrieval targets for generating evaluation QA datasets.

Description

Passage Sampling selects a subset of passages from a corpus to serve as retrieval ground truth (retrieval_gt) for QA generation. Each sampled passage becomes the basis for generating a question-answer pair. The sampling strategy determines the structure of the resulting QA dataset: single-hop sampling selects individual passages (one passage per question), while multi-hop sampling selects groups of passages (requiring synthesis across multiple passages to answer). Random sampling ensures diversity across the corpus.

Usage

Use this principle after corpus construction and before query generation. It determines which passages will have questions generated about them. Random single-hop is the default for simple factoid QA datasets; use multi-hop sampling when you need to evaluate cross-passage reasoning capabilities.

Theoretical Basis

The sampling process creates a mapping from questions to ground-truth passages:

Single-hop sampling:

Select N passages uniformly at random from the corpus
Each passage becomes one retrieval_gt entry: $r e t r i e v a l_g t_{i} = [[p a s s a g e_i d_{i}]]$
Resulting QA pairs test single-passage retrieval

Multi-hop sampling:

Select groups of K passages that are related
Each group becomes one retrieval_gt entry: $r e t r i e v a l_g t_{i} = [[i d_{1}, i d_{2}, . . ., i d_{K}]]$
Resulting QA pairs test multi-passage retrieval and synthesis

Related Pages

Implemented By

Implementation:Marker_Inc_Korea_AutoRAG_Corpus_Sample_Random_Single_Hop

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment