AI Summarization — AI Glossary

Reference details

Related terms

• Document Q&A
• ROUGE
• Long-Context Model

How it works

1. Short inputs are summarised in one call, which preserves global context because the model sees everything at once.
2. Long inputs exceed the window, so they are chunked and summarised per chunk, then the summaries are summarised — the map-reduce shape.
3. That shape loses cross-chunk relationships, because no call ever sees two distant parts of the document together.
4. Hierarchical approaches build a tree of summaries so higher levels retain structure the flat reduce step would discard.
5. Query-focused summarisation conditions on what the reader wants, which changes what counts as important and usually beats a generic summary for a specific need.

Example: What map-reduce cannot see

A 200-page report split into 20 chunks of 10 pages. A contradiction between page 8 and page 180 sits in chunks 1 and 18, and no summarisation call ever receives both — so the reduce step combines two summaries that each omitted it as locally unremarkable. The output is fluent, accurate per chunk, and silent on the most important finding. Hierarchical summarisation mitigates this by keeping intermediate levels; a single call over the whole document avoids it entirely when the window allows.

Commonly confused with

Treated as one task when it is several with different failure modes. Extractive summarisation selects existing sentences and cannot misstate the source, but reads poorly; abstractive summarisation rewrites and can introduce claims the source never made. Single-document and multi-document are different again — the latter must reconcile disagreement between sources rather than condense one. And compression is not the goal: a summary that drops the one finding that mattered is shorter and worse, which is why query-focused variants exist.

When to use it

Reach for it when:

• Long content must be condensed for a reader who will act on the summary.
• The window cannot hold the document, making a hierarchical or map-reduce approach necessary.
• The reader's question is known, where query-focused summarisation outperforms a generic one.

Reach for something else when:

• Specific facts are needed, where retrieval answers directly without compressing away the rest.
• Cross-document contradictions matter and a flat map-reduce would hide them.
• Verbatim fidelity is required, where abstractive rewriting is a risk rather than a feature.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Summarization — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path