query_id stringlengths 4 4 | query stringlengths 52 82 | relevant_ids listlengths 1 1 |
|---|---|---|
q_01 | How does RRF combine results from different rankers? | [
"doc_01"
] |
q_02 | What makes cross-encoders more accurate than comparing two separate embeddings? | [
"doc_02"
] |
q_03 | Explain how keyword frequency scoring works in sparse search | [
"doc_03"
] |
q_04 | How do neural embeddings enable similarity search between text? | [
"doc_04"
] |
q_05 | How are real-world facts represented so a system can traverse them? | [
"doc_05"
] |
q_06 | What is late chunking and why keep document-level context in each chunk? | [
"doc_06"
] |
q_07 | How does parent-child chunking balance precise matching with full context? | [
"doc_07"
] |
q_08 | How do you check if a generated answer is actually grounded in what was retrieved? | [
"doc_08"
] |
q_09 | What metric caps answer quality based on how much needed info was retrieved? | [
"doc_09"
] |
q_10 | Why would a system generate a fake answer before embedding a search query? | [
"doc_10"
] |
q_11 | How does rephrasing one question into several versions help retrieval? | [
"doc_11"
] |
q_12 | What is the benefit of asking a broader question before the specific one? | [
"doc_12"
] |
q_13 | How does a system decide which retrieval approach to use for a given question? | [
"doc_13"
] |
q_14 | How do you split text based on topic shifts instead of a fixed length? | [
"doc_14"
] |
q_15 | What step connects a mentioned name to a node in a graph? | [
"doc_15"
] |
q_16 | What tradeoff do fast approximate vector search methods make? | [
"doc_16"
] |
q_17 | How do you measure whether a generated answer stays on topic? | [
"doc_17"
] |
q_18 | What is the simplest way to break a document into pieces? | [
"doc_18"
] |
q_19 | How can questions requiring multiple connected facts be answered? | [
"doc_19"
] |
q_20 | How can a piece of text's embedding reflect the context around it? | [
"doc_20"
] |
RAGBench Queries
A collection of evaluation queries designed for benchmarking Retrieval-Augmented Generation (RAG) systems.
Dataset Description
RAGBench Queries contains test queries used to evaluate different retrieval and chunking strategies in a RAG pipeline.
The dataset was created as part of the RAGBench project, which compares retrieval performance using different document chunking approaches.
Purpose
The queries are designed to evaluate whether a RAG system can retrieve the correct information from a collection of documents.
The benchmark can be used to compare:
- Semantic chunking
- Parent-child chunking
- Other retrieval strategies
- Different embedding models
- RAG pipelines and retrieval configurations
Dataset Structure
The dataset is provided in JSON format.
Each query contains information required to evaluate retrieval performance, including the query and its expected relevant document.
Example:
{
"query": "Example question about the document",
"relevant_doc": "document_1"
}
- Downloads last month
- 33