The licensing trap in RAG data

Most teams building a Retrieval-Augmented Generation (RAG) system, where a language model answers questions using documents you retrieve at query time, spend weeks worrying about chunk sizes and embedding models. Then they wire in a dataset, ship the thing, and only later stop to read the licence. That is the trap. The data was the easy part to find and the easy part to overlook.

We are not here to scare you. Licensing is rarely a disaster, and most of the good open datasets are genuinely usable. But a licence is a contract, and a few common clauses can quietly reshape what you are allowed to build. It is worth knowing which ones to look for before they surprise you.

Share-alike can follow your data around

Some of the most popular open datasets carry a share-alike clause. In plain terms, that means if you build on the data and share the result, you have to release your result under the same licence. Wikipedia’s text is the classic example: it is licensed under CC BY-SA, which asks you to attribute the source and pass on the same freedoms.

For a lot of RAG systems this is fine, because you are retrieving passages to answer a question, not republishing the whole corpus. But the moment you start distributing something derived from that data, an enriched dataset, a fine-tuned model, a bundle of pre-computed embeddings, the share-alike obligation can travel with it. The question to ask is not “can I use this?” but “if I share what I make, what am I agreeing to share with it?”

Non-commercial means what it says

A licence marked non-commercial (often CC BY-NC) lets you use the data for research, learning, and personal projects, but not to make money. That sounds clear until you are the one deciding whether an internal tool at a company counts as commercial use. It usually does.

This one catches teams because the data is free to download, so it feels free to use. Free to access and free to build a business on are two different things. If you are prototyping, a non-commercial dataset is a fine place to learn. If that prototype is heading for production behind a paywall or inside a paying customer’s workflow, you will want a dataset whose terms allow it, or a commercial licence from the people who made it.

”Open” is not always as open as it looks

The word “open” does a lot of heavy lifting, and it is not a legal term. A dataset can be described as open and still come with real conditions: attribution requirements, restrictions on redistribution, terms that only cover the collection and not the underlying content, or a sign-up step that quietly binds you to a longer agreement.

Web-crawl corpora are a good example. The dataset itself may be openly licensed, but the individual pages inside it are still owned by whoever published them, and their terms have not gone away. “Open” is a starting point for a conversation, not the end of one. Read past the label to the actual licence text, and note who owns the content as well as who compiled it.

What to do about it

None of this needs to slow you down. A short habit at the start of a project saves the awkward conversation later:

• Find the actual licence, not a summary, and read the specific clauses on commercial use, redistribution, and share-alike.
• Separate the licence on the dataset from the ownership of the content inside it.
• Ask where your project is heading. Internal demo, public product, and paid service each carry different requirements.
• When the terms are unclear, treat that as a reason to check with the maintainer or a lawyer, not a reason to assume the friendly answer.

We try to make this easier by listing the licence for every source in our directory, so you can see the terms before you commit to a dataset. If you want the concepts behind the labels, our learn pages go deeper on licences, formats, and RAG readiness. And if you are curious about why we care enough to catalogue all this in the first place, that is the story on our about page.

Data licensing is not the exciting part of building with RAG. But an hour spent reading terms early is a lot cheaper than rebuilding on a different corpus later, and it lets you ship with a clear head.