fenxer/macos-summarize-internals

When you select text on a Mac and right-click to choose "Summarize," macOS gives you back a shorter version of that text. This repository figures out exactly how that works under the hood. The author reverse-engineered the system component called SKSummary, which lives inside Apple's SearchKit framework, and then rebuilt the same algorithm in Python. The Python version produces output that matches the real macOS tool almost perfectly across a set of 12 test files, with 9 byte-identical matches and 2 others differing only in tiny floating-point noise. The big finding is that macOS summarization is not a neural network or any modern machine learning model. It is an old-school search engine technique. The system splits your text into sentences, builds a search index over them, then treats the entire document as a search query against that index. Each sentence gets a score based on how similar it is to the whole document, and the top-scoring sentences become the summary. The default summary length is based on a logarithmic formula, and the slider in the UI simply switches to a percentage of total sentences. The README documents many specific quirks the author discovered through probing. There is no stopword removal and no stemming. Numbers like "3.5" get split into "3" and "5" at the decimal point. Email addresses split at the @ symbol but the domain stays intact. Sentence segmentation is particular about periods followed by lowercase letters, and Chinese punctuation marks always trigger a break. When sentences tie in score, the later one ranks higher. The repository includes the Python reimplementation, C programs that call the real macOS system API to dump ground-truth data, a comparison script that checks the Python version against the system, and a folder of probes and experiments from the reverse-engineering process. A full write-up in both English and Chinese tells the story from disassembly through verification. The full README is longer than what was shown.