Summarize documents that do not fit in context
Even large context windows have limits, and feeding a 200-page report to a single call is slow and expensive. Recursive summarization solves this with a map-reduce approach: summarize each chunk, then summarize those summaries, repeating until one coherent summary remains. This tool does it in your browser with your own OpenAI or Anthropic API key.
How it works
You paste the document, set a chunk size, and provide your key. The tool splits the text on paragraph and sentence boundaries into chunks of roughly your chosen size, then summarizes each chunk with a real API call. It concatenates the chunk summaries and, if they are still long, summarizes them again — repeating this reduction until a single summary is produced. Progress is shown as each layer completes, and the final summary plus the intermediate chunk summaries are displayed. Every call goes directly from your browser to the provider.
How many passes does it take?
The number of reduction passes depends on the document size and your chunk size. For a 100,000-character document split into 6,000-character chunks, pass 1 produces about 17 chunk summaries. If those summaries together still exceed the chunk size, pass 2 summarizes them in turn — typically yielding just 2 or 3 summary-of-summaries. Pass 3 (if needed) collapses those into the final single paragraph. Most documents reach a final summary in 2–3 passes, regardless of original length.
Illustrative example
Suppose you have a 40-page technical report (~80,000 characters). At a 6,000-character chunk size:
- Pass 1: ~14 chunk summaries produced (roughly one paragraph each)
- Pass 2: Those 14 summaries are themselves summarized into 2–3 paragraphs
- Pass 3: The 2–3 paragraphs collapse into one final executive summary
The whole run makes roughly 16–18 API calls. At a modest per-call cost, this is far cheaper than a single call on a very large context and also faster when chunk calls run in parallel.
What you gain vs. lose
Recursive summarization excels at extracting the main conclusions, themes, and structure of a document. What it compresses away is fine-grained detail — exact figures buried in appendices, specific clause wording in contracts, and low-salience paragraphs will not survive many reduction passes. This makes it ideal for: executive summaries of lengthy reports, getting a quick orientation to an unfamiliar document before reading it carefully, and generating topic-level questions to ask with more targeted retrieval later.
Tips for better summaries
- Tune chunk size to the trade-off. Larger chunks preserve more context per summary but cost more tokens per call; smaller chunks are cheaper per call but lose cross-chunk context. A 4,000–8,000 character range works well for most documents.
- Use a cheap model for the first pass. GPT-4o mini or Claude Haiku handle chunk summarization well; reserve a stronger model for the final reduction if you need a polished result.
- Clean the input first. Stripping boilerplate, repetitive headers, footers, and page numbers before summarizing improves quality and meaningfully cuts cost.
- Write a directive prompt. Instead of a generic “summarize this”, the tool lets you specify a focus — for example, “summarize the methodology and key findings only” — which each chunk call inherits.
- Expect some detail loss in deep passes — that is inherent to compression.