Ref Long Documents 50pp

Reference — Long Documents (50+ Pages)

Scope

Any legal document exceeding approximately 50 pages — a long-form contract, an M&A due diligence data room, a regulatory submission, a multi-day arbitration hearing bundle, a legislative comment package — requires a different processing strategy from short documents. Pasting the full text into a prompt has three problems:

1. Token cost: Long documents consume a disproportionate share of the context window, leaving little room for the AI to reason and respond
2. Attention diffusion: AI models do not read every word with equal attention across very long contexts; material near the middle of a very long prompt is more likely to be mischaracterized or overlooked
3. No targeting: Pasting the full text and asking "summarize this" produces a generic summary; targeted analysis of specific clauses, sections, or risk areas requires targeted extraction

This reference sets out the strategies for handling long documents efficiently.

Strategy 1 — Outline first, then targeted analysis

The single most effective approach for large documents:

Step 1 — Extract the structure Prompt the AI to read the document (or a chunk if it is very long) and produce only a structural outline:

• For a contract: article numbers, headings, and one-sentence summary per article
• For a judgment: procedural history, issues, findings, and outcome — without detail
• For a data room: document names, dates, and one-line description per document

Step 2 — Identify priority sections From the outline, identify which sections require detailed analysis. In an M&A context, this might be the representations and warranties, the material adverse change definition, and the closing conditions — not the boilerplate.

Step 3 — Targeted analysis per section Analyze each priority section in isolation: paste only that section, with full context about what it is, and apply the relevant review skill.

This approach produces better output than a single "analyze everything" prompt because each targeted analysis has the full context window available for reasoning.

Strategy 2 — Chunked processing with rolling summary

For very long documents where even the outline is too long to process in one pass:

Process:

1. Divide the document into chunks of approximately 5,000–8,000 words (roughly 15–25 pages)
2. For each chunk: produce (a) a summary of key provisions and (b) a list of flagged issues
3. At the end of all chunks: consolidate the summaries and flag lists into a master document

Key rule: Each chunk summary must include enough context to stand alone — the consolidated summary must make sense without re-reading the full document.

Variant — Rolling context: Pass the summary of chunk N as context when processing chunk N+1. This allows the AI to detect cross-chunk issues (e.g., a definition in Article 1 that affects a clause in Article 47).

Strategy 3 — Scanned PDF extraction

Many long legal documents arrive as scanned PDFs (executed contracts, court files, physical records). These cannot be processed as text until the text is extracted.

Apply [[multimodal-scanned-pdf-handler]] for:

• OCR extraction of scanned PDFs
• Handling dual-column Arabic / English layouts
• Handling handwritten notes or stamps on official documents
• Quality-checking the extraction before analysis

Quality threshold: If OCR confidence is below 90%, the extracted text may contain errors that corrupt the analysis. Flag low-confidence extractions and prompt the user to provide a cleaner source.

Strategy 4 — Vector embedding for semantic search

For very large document sets (a data room with 200+ documents) where even chunked processing of each document is impractical:

Approach:

1. Generate vector embeddings for each document chunk (using the platform's embedding model)
2. When a specific question arises ("does any document contain a change-of-control provision?"), run a semantic search against the embedding store to retrieve the most relevant chunks
3. Analyze only the retrieved chunks, not the full document set

Applications in legal work:

• Due diligence: "Does any contract in this data room contain a most-favored-nation clause?"
• Litigation discovery: "Which of these 500 emails discusses the project delivery date?"
• Regulatory audit: "Which of these policies reference the PDPL's data subject rights obligations?"

This approach converts a document review task from a linear sequential process into a query-driven one — dramatically more efficient for large sets.

Strategy 5 — Agent tool-call patterns for document navigation

For platform implementations with tool-calling (MCP) enabled:

Preferred pattern:

1. Tool call: load_document(path) — indexes the document and returns a structural outline
2. Tool call: navigate_to_section(article_id) — retrieves the specific section text
3. Tool call: analyze_clause(section_text, analysis_type) — runs the appropriate review skill on the isolated clause
4. Tool call: compile_findings() — assembles the section-level findings into a master report

This pattern is more reliable than a single large prompt because:

• Each tool call is atomic and independently verifiable
• The agent can navigate back to earlier sections if a later clause depends on an earlier definition
• The output is structured and traceable to specific document sections

Anti-patterns to avoid

Decision guide

Related skills

• [[ref-anti-patterns]]
• [[ref-verification]]
• [[multimodal-scanned-pdf-handler]]
• [[public-tool-contract-summarizer-public]]
• [[public-tool-case-summarizer-public]]