Text and Data Mining (TDM)

Last updated: 05/10/2026

How we respect publishers' rights reservation

Kevra has a feature called the reading companion: a guide to each book's characters, places and events that only reveals as much as each reader has read. To build it, the text of the chapters is processed by a language model once per title.

We respect the TDM opt-out

Before processing a book, we check whether the work reserves its text and data mining rights. If the reservation is set, we do not process that book: its reading companion is not built, then or later, and the title is flagged so that it is not built from someone else's copy either.

We follow the TDMRep protocol from the W3C Community Group (final report of 10 May 2024), the machine-readable way of exercising the reservation provided for in Article 4(3) of Directive (EU) 2019/790. In an EPUB we read the metadata of the package document:

<meta property="tdm:reservation">1</meta>
<meta property="tdm:policy">https://publisher.example/policy.json</meta>

We also read the older EPUB 2 form (<meta name="tdm:reservation" content="1"/>). The value 1 means rights reserved. If a policy is declared in tdm:policy, we record it: it is the door to knock on should we ever want to ask for permission.

What we do not do

• We do not train models on anyone's books. The text is used to build the guide for that work and is not kept afterwards.
• We do not process DRM-protected books. No decrypted text ever leaves the device from a publication protected with LCP.
• We do not use the text of a reserved book by an indirect route —for example, from another copy of the same title without the flag. The decision is made per work.

If you are a publisher and want to reserve them

Including the flag in your files is enough: we will read it on our own, without you having to write to us. If you would rather tell us directly, or want to lift a reservation that no longer applies, write to derechos@joinkevra.com.