TractorBeam

TractorBeam is a browser extension that allows users to define, iterate on, and evaluate AI highlighting systems on PDFs. These highlight systems can serve a variety of purposes — reading support and document indexing, acting like a sort of semantic ctrl-F; rapid prototyping for information extraction systems, e.g. extracting all instances of names, materials, or abstract concepts from a document; and adversarial refinement of your own understanding of a domain: updating your mental model based on the differences between how you think of it and how an AI highlights it.

Background

TractorBeam takes the form of a collaborative annotation tool: a tool that allows a group of users to jointly highlight the contents of a document. Systems like this (including Hypothesis and Genius) allow shared meaning making through these highlights: people can see what other people found interesting, notes and comments can facilitate discussion, and the highlights are direct responses to the text that remain grounded to that text: users are reading comments alongside the text that originated them.

These properties of collaborative annotation are also really useful for AI: hallucinations are less of a concern if you’re seeing an AI’s results in light of the original document, and evaluating whether an AI is doing a good job is a lot easier when you see what it’s drawing from. I wrote about some of the benefits of collaborative annotation for machine learning in a paper at the Science and Technology of Augmented Reading Workshop at CHI 2026, and you can find my paper here. TractorBeam is also forthcoming as a poster at UIST 2026 (paper|poster), and a full version of the study is now under review.

While TractorBeam doesn’t support collaboration with other users yet, it frames AI as a collaborative annotator on the same documents you read, in a very common, existing PDF reader (Mozilla’s pdf.js). The AI highlights are clearly marked as separate from yours, and you can choose to use them to either better reflect your mental model in instances, or to refine the model’s behavior and re-run it.

Try it out!

If you’d like to try TractorBeam, a (poorly documented) beta version is available in the Chrome Web Store here. It’s also open-source, and available on GitHub.

If you have any questions, please shoot me an email! I’m happy to help you get started with TractorBeam.

TractorBeam’s privacy policy is available here.