Learning path · Multimodal · 84
Summarization Patterns
Map-reduce, hierarchical, and extractive-abstractive blends for long content—docs, calls, threads.
Why it matters
- Core enterprise use case with measurable ROI.
- Pattern choice affects faithfulness on long inputs.
- Pairs with eval rubrics for omission and distortion.
Key ideas
- Map-reduce
- Refine loops
- Structured summaries
Top resources
- 01DocsOpenAI
Prompt engineering
Why this resource. Map-reduce and structured summaries start as prompt/program design.
Covers in this concept
- map-reduce
- structured output
- 02DocsOpenAI
Images and vision
Why this resource. When the source is a document image or slide, not a clean text file.
Covers in this concept
- documents
Summarization patterns manage length: map-reduce summarizes chunks then merges; refine iteratively updates a running summary; structured outputs force sections—decisions, risks, action items. Pick patterns based on fidelity needs—legal summaries may require extractive anchors. Evaluate with G-Eval style rubrics for omission, not only fluency. Require structured sections—decisions, risks, owners—in executive summaries so readers can skim reliably under time pressure. Compare extractive anchors against abstractive prose on compliance-sensitive summaries.
Updated 2026-08-09 · Full learning path