Key answer
Choose according to information needed: full text preserves the most context, excerpts preserve local wording, and summaries convey the main thread. Supply original passages when conditions, exceptions or reasoning steps affect the answer.
Full explanation
| Form | Suitable tasks | Information at risk |
|---|---|---|
| Full text | Overall structure, cross-chapter contradictions, complete rules and exceptions | Parsing or current retrieval may still cover only part |
| Excerpt | Rewriting a passage, checking a claim or answering locally | Surrounding definitions, scope and exceptions elsewhere |
| Summary | Orientation, initial discussion and handoff | Exact wording, detail, evidence chains, unusual cases and unresolved disagreements |
Also check summary accuracy. Maynez et al.'s ACL 2020 human evaluation of summarization systems found source-inconsistent content. Although about models of that period, it illustrates errors introduced by rewriting as well as information lost through compression. Original study
More full text is not automatically better. *Lost in the Middle* shows continuing difficulty finding and using relevant long-context information. For long documents, provide contents or a short guide, state the question and supply key sections; for global omissions, cover the complete contents systematically. Original study
Here are the contents and summary for orientation. This task concerns whether Section 4's conclusion holds; its original text, tables and appendix conditions are attached. Identify other sections needed for judgment.
Attach titles, pages and versions to excerpts, retaining definitions and exceptions. With summaries alone, limit the claim to a preliminary summary-based judgment, not verification of original wording, all evidence or absence of counterexamples throughout the text.
References
1. Maynez et al. (2020), On Faithfulness and Factuality in Abstractive Summarization, ACL
2. Liu et al. (2024), Lost in the Middle: How Language Models Use Long Contexts, TACL