Automatic Text Summarization

Long documents are slow to read, and not every sentence in one carries equal weight: most of a document's meaning tends to be concentrated in a handful of central sentences, surrounded by supporting detail, examples and asides. Skimming a full document just to find its main point doesn't scale once there are hundreds of documents to get through, which is the same problem retrieval in Chapter 3 solves for finding documents in the first place. This chapter is about condensing a single one, once you're already looking at it.

5.1 Two Families of Summarization

Summarization comes in two families, and the toggle above switches between working examples of both. Extractive summarization selects a handful of the document's own sentences, verbatim, and presents them in their original order as the summary: nothing is paraphrased or invented, which makes it far easier to trust, since there's no way for it to introduce a claim the source document never made, but it can only ever rearrange sentences that already exist. Abstractive summarization asks a language model to write a new, shorter version of a text in its own words, the same generation mechanism from Chapter 2, given a document instead of a question: it can compress across sentences, merge related ideas and use its own phrasing, at the cost of a real model call and the small chance it paraphrases something in a way the source didn't quite mean.

5.2 How Each Mode Works

Abstractive mode simply hands your text to a language model with an instruction to paraphrase it in roughly your target length, without copying sentences verbatim or adding claims the source doesn't make, the same generation step Chapter 2 uses for chat. If a live model connection isn't available, the tool clearly labels its reply as a placeholder rather than a real paraphrase, since a genuine paraphrase requires an actual model call to produce. Extractive mode instead uses TextRank, a graph-based ranking algorithm: each sentence is embedded (the same technique from Chapter 3) and compared against every other sentence with cosine similarity, turning the document into a graph where a sentence's score depends on how well-connected it is to the rest of the document's ideas, the same random-walk idea behind PageRank, applied to sentences instead of web pages. Unlike abstractive mode, this runs for free, with no model call involved at all.

5.3 Two Smaller, Older Techniques

Extractive mode also runs two much older and simpler techniques over the same text. Keyword extraction counts how often each word appears after common function words like "the" and "of" are filtered out; the surviving high-frequency words are a decent guess at what a passage is actually about. It's the "term frequency" half of TF-IDF, a term-weighting idea from information retrieval that predates embeddings by decades, included here to show that not every useful text technique needs a model at all.

Readability scoring answers a different question: not what the text is about, but how hard it is to read. The Flesch-Kincaid formulas turn three simple counts (words, sentences and syllables per word) into an estimated U.S. school grade level. Run it on both the original text and either mode's summary and compare: extractive summaries often score a lower (easier) grade level than the original, purely because TextRank tends to favor shorter, self-contained sentences, even though not a single word was rewritten to simplify it, while an abstractive summary's grade level depends entirely on how the model chose to phrase its paraphrase.

5.4 Why Not Just Always Use the LLM?

Given a capable model is already one prompt away, it's worth asking why extractive methods are still used in production systems at all, instead of every summarizer just calling an LLM. Two practical reasons: cost and verifiability. An abstractive call bills real tokens on every request, the same way Chapter 1's tokenizer counts and Chapter 2's chat does, while TextRank's embedding-similarity pass, like Chapter 3's retrieval or Chapter 6's cache, runs for free. And because an extractive summary only ever rearranges the source's own sentences, it can never introduce a claim the document didn't make, exactly the hallucination risk Chapter 2 covers, a much stronger trust guarantee than any abstractive rewrite can offer. Reach for extractive when the source needs to be quoted exactly or cost matters at scale; reach for abstractive when a compact, readable paraphrase matters more than either.