@document_analysis/long_pdf_extraction
multi-needle retrieval is the capability cliff for Sonnet 4.6
I need to extract key facts from a 300-page technical report
locate@document_analysis/long_pdf_extraction
Opus 4.8 for multi-section cross-reference and citation precision. Sonnet 4.6 for single-pass structured extraction at lower cost. All outputs require source verification regardless of model.
structured PDF, single extraction pass, cost-sensitive
claude/sonnet-4.6multi-section cross-reference, citation accuracy required, >100 pages
claude/opus-4.8output will be used without manual verification
openai/gpt-5.5any model output used for citation
human| ID | Relation | Actor | Result | Output |
|---|---|---|---|---|
| CT-D4-001 | locate@cross_section | claude/opus-4.8 | accepted | strong multi-needle retrieval; improved citation precision; cheaper token cost in workflow tests |
| CT-D4-002 | locate@cross_section | claude/sonnet-4.6 | bounded | matches Opus on single-pass structured docs; degrades on multi-needle retrieval |
| CT-D4-003 | validate@citation | openai/gpt-5.5 | rejected | high hallucination rate makes unverified extraction unreliable |
Opus 4.8 on 8-needle 1M MRCR v2: 76%; Sonnet baseline: 18.5%
E-GLOBAL-010Sonnet 4.6 matches Opus 4.6 on OfficeQA enterprise document reading
E-GLOBAL-005GPT-5.5 hallucination rate 86% on AA-Omniscience
E-GLOBAL-003