How to choose your first RAG dataset
Practical, friendly advice for picking your first RAG dataset: start small, check the licence and format, and match the data to what you are building.
Picking your first dataset for a Retrieval-Augmented Generation (RAG) system, meaning an app that searches a pile of documents and quotes from them to answer questions, can feel like standing in front of a very large buffet. There is so much data out there, and it is tempting to grab the biggest, most impressive-sounding corpus you can find. Resist that urge. Your first dataset is for learning how the pieces fit together, and a smaller, cleaner one will teach you far more, far faster.
Here is how we would go about it.
Start small and cleanly licensed
The best first dataset is one you can download in minutes and reason about easily. A few thousand documents is plenty to build a working prototype, see retrieval in action, and spot where things go wrong. A dataset measured in terabytes will just slow you down while you learn.
Wikipedia is our usual recommendation. It is broad, well structured, and comes in tidy formats. You can grab a single-language dump, or even a themed subset, and have something meaningful to search within an afternoon. Because the content is familiar, you will immediately notice when your system retrieves the wrong passage, which is exactly the feedback you want early on.
Cleanly licensed matters just as much as small. A licence is the written permission that says what you are allowed to do with the data: store it, index it, show snippets to your users, and build a product on top. Starting with openly licensed data means you never have to unpick a licensing problem later, when you are more invested. You can read more about the licences you will meet in our guides, but the short version is: for a first project, stick to open, permissive sources.
Match the data to your use case
A dataset is only good if it holds the answers your users will actually ask for. This sounds obvious, but it is the step people skip most often.
Write down three or four real questions you want your system to answer. Then ask whether your candidate dataset genuinely contains that knowledge. A general encyclopaedia is wonderful for broad factual questions, but it will not help a tool that answers questions about your company’s internal policies, your local council’s planning rules, or the latest research in a niche field. If your use case is specialised, you may need a specialised source, or your own documents, rather than a big general corpus.
When you are ready to move beyond Wikipedia, our directory groups sources by subject, from legal and biomedical to code and news, so you can find data that fits what you are building rather than the other way round.
Check the licence and the format
Two practical checks will save you a lot of grief.
First, the licence, as above: confirm you are allowed to use it for whatever you have in mind, especially if there is any chance of money changing hands one day.
Second, the format. Data arrives in many shapes: plain text, JSON, Parquet, CSV, XML, and web-archive formats like WARC. Some are a joy to work with; others need real effort to unpack before you can use a single sentence. Before you commit, download a small sample and check you can actually open it and pull clean text out. A dataset you cannot parse is not a dataset you can use yet.
Mind the RAG-readiness
Not all data is equally ready to drop into a RAG pipeline. We think of it as a spectrum.
Raw data is exactly as it was collected: messy, full of navigation menus, adverts, and boilerplate. Cleaned data has had that clutter stripped out. Further along, some datasets are already split into passages (chunked), and a few even come as embeddings, meaning the text has been converted into the numerical form a vector database searches through, so you can plug it straight in.
For your first project, cleaned data is the sweet spot. Raw data means you will spend your time on tidying rather than learning how retrieval behaves. Pre-chunked or embedded data can be brilliant later, but it hides steps you probably want to understand first.
Put it all together
Choosing well comes down to a few friendly habits: start small, pick something openly licensed, make sure the data actually contains the answers you need, check you can open the format, and favour data that is reasonably clean. Get a modest Wikipedia-based prototype working end to end, and you will understand RAG far better than if you had wrestled a giant corpus for a week.
When you are ready for your second dataset, the directory is waiting, and you can always read more about what we do and why we built it.