RAG-250: Multi-Hop Retrieval Benchmark

• The paper introduces Multi-Meta-RAG, a method that enhances standard RAG pipelines by using lightweight metadata extraction and filtering to improve retrieval quality.
• RAG-250 is a multi-hop retrieval benchmark requiring evidence chaining from multiple documents with constraints on source and publication date.
• The methodology leverages database filtering and semantic re-ranking, demonstrating substantial improvements in answer accuracy on related multi-hop QA workloads.

RAG-250 can be thought of as a multi-hop retrieval-augmented generation benchmark with around 250 questions that require chaining evidence across multiple documents and, in many cases, respecting constraints about sources and dates. In the literature summarized here, the clearest method designed for this setting is Multi-Meta-RAG, which adds a lightweight metadata extraction step to the front of a standard RAG pipeline and then uses database filtering to eliminate irrelevant candidates before semantic retrieval and re-ranking. On MultiHop-RAG—the benchmark on which the method was developed—this addition substantially improves retrieval quality and final answer accuracy, and the same gains are described as transferring naturally to a RAG-250 workload (Poliakov et al., 2024).

1. Definition and task profile

RAG-250 is characterized by multi-hop retrieval and reasoning requirements. Its questions often name specific sources and may include temporal constraints. A representative form is a comparison such as whether one outlet reported an event before another outlet reported a different event. Answering such queries requires collecting multiple pieces of evidence from the right outlets and dates, then chaining them through bridge entities, topic links, comparisons, or temporal ordering (Poliakov et al., 2024).

This setting differs materially from single-document factual lookup. The retrieval problem is not merely to find semantically related passages; it is to recover a constrained evidence set whose members satisfy explicit or implicit predicates in the query. In the RAG-250 description, the minimal metadata fields are source and published_at, because the questions often hinge on outlet identity and publication date. A plausible implication is that RAG-250 belongs to the class of constrained multi-hop QA workloads in which retrieval must jointly satisfy topical relevance and metadata compliance (Poliakov et al., 2024).

The workload also