Building an AI summarisation tool
An AI summarisation tool takes long content — an article, a report, a meeting transcript — and returns a short, faithful summary. The hard parts are not the model call itself but everything around it: cleaning messy input, fitting long text into the context window, and forcing a consistent output structure. This guide walks through the full pipeline, and the builder below assembles the exact summarisation prompt for a piece of text you paste in.
How the pipeline works
The tool has four stages. Ingest accepts pasted text or a URL; for a URL you fetch the page server-side and run a readability pass to keep only the article body. Chunk checks whether the cleaned text fits the model’s context — if not, it splits the text into overlapping passages. Summarise runs a map-reduce pass: each chunk is summarised, then the chunk summaries are summarised together into one final result. Format asks the model for a fixed structure so the output is always a TL;DR plus key points.
The map-reduce step is what lets the tool handle documents far larger than any single context window. Short inputs skip straight to a single summarisation call; long inputs fan out and then converge.
Tips and cost notes
Specify the output shape explicitly — a one-sentence TL;DR, three to five bullet points, and a hard length cap — so summaries are predictable. Strip boilerplate before sending, because every wasted token costs money and dilutes focus. Cache summaries keyed by a hash of the source so repeat requests are free and instant. For long documents, keep a small overlap between chunks so no sentence is split across a boundary and lost. Pick a smaller model first; upgrade only where nuance demands it.