New AI Maps Inventions to Make Patent Searches Faster and Cheaper

Every patent is basically a long, legalistic description of how an invention works, and many run past 40,000 words — longer than a short novel. Today's AI search tools usually cut those documents off after a few thousand words, the way you might skim only the first chapter of a book before deciding it isn't what you need. That means they miss details buried near the end, which is often exactly where the important claims live.

Two researchers, Nikolai Zenovkin and Sebastian Björkqvist, took a different approach. Instead of scanning raw text, their system draws a map of each invention — an "invention graph" that links the pieces of an idea (a battery, a circuit, a coating) and shows how they connect. Think of it as turning a rambling description into a labelled diagram you can actually navigate. An earlier version of this mapping was done by hand-written rules, which broke easily; the new version learns the patterns itself.

The clever part is how it was built. It learned by studying 1 million patents already mapped out by that older rule-based software, essentially copying a teacher it later outperformed — while running three times cheaper. Because it scores text in small sliding windows rather than comparing every word to every other word, it can train on short snippets and still handle full-length documents without being retrained.

Why should you care? Patents shape drug prices, gadget features, and who gets sued. Faster, cheaper, more accurate searches help small inventors check their ideas without paying lawyers thousands, help backlogged patent offices move quicker, and make it harder to sneak through weak patents used mainly for lawsuits. The catch is that the improvement is real but modest — roughly 0.5% to 1.1% better at finding relevant citations — and this is research, not a product you can use tomorrow.