MIRACL
MIRACL (Multilingual Information Retrieval Across a Continuum of Languages) is an evaluation benchmark, not a knowledge base. It measures monolingual retrieval, where the query and the documents share a language, across 18 languages including Arabic, Bengali, Chinese, Hindi, Russian, Swahili, and Yoruba. Native-speaker annotators judged which Wikipedia passages answer each query, producing hundreds of thousands of relevance labels and a fair basis for comparing retrievers.
Everything you need is on Hugging Face: the per-language passage collections already split into retrieval-sized chunks, the query topics, and the relevance judgements (the qrels). Because the corpus is pre-chunked, the workflow is short: load a language, embed its passages into a vector database, run the queries, and score the results against the human labels with standard metrics such as nDCG and recall.
It has become a go-to reference for anyone building retrieval for non-English audiences, where results from English-only benchmarks often fail to carry over. Use it to choose an embedding model, tune your chunking strategy, or confirm your pipeline genuinely works in the languages your users actually speak.
Keep its purpose in view. This is a test set, not production content: the passages are a fixed Wikipedia snapshot, so do not serve MIRACL as a live knowledge base. Coverage per language is uneven, and a strong score here proves competence on encyclopaedic text, not on your own domain. The benchmark is released under Apache 2.0, which makes it straightforward to use in commercial evaluation work, though the passages themselves originate from Wikipedia.
It complements the English-focused MS MARCO and Natural Questions, and pairs well with mMARCO and the wider BEIR suite when you want a fuller picture of multilingual and cross-domain retrieval quality.
Related sources
BEIR
A collection of 18 information retrieval datasets spanning ad-hoc web search, question answering, fact verification, and duplicate question retrieval. Built by aggregating existing datasets, some originally created for other tasks and converted to retrieval format. Tests how well retrieval models generalise to unseen domains without fine-tuning, which matters for real-world RAG.
BRIGHT
A reasoning-intensive retrieval benchmark. It requires genuine reasoning to connect queries with relevant documents rather than keyword or semantic overlap, revealing model weaknesses that BEIR misses.
CoIR (Code Retrieval Benchmark)
A code retrieval benchmark spanning diverse programming tasks and languages. Built for evaluating how well models retrieve relevant source code, which matters if you are building RAG over codebases.
FreshStack
A 2025 framework and benchmark for retrieval over fast-moving technical documentation. It pairs real Stack Overflow questions and answers with chunked GitHub code and docs across five niche domains, giving code and docs RAG a hard, contamination-resistant test.