Contract Clause Extraction Pipeline

Gemini 2.5 Pro handles multi-column, table-heavy, and scanned contracts with page-coordinate grounding. Unstructured.io is cheaper for native PDFs but misses scanned contracts. Textract is solid for tables but needs LLM cleanup for legal formatting.

Alternatives: Claude Sonnet 4 vision, GPT-4o vision, Unstructured.io, Textract + cleanup

Sonnet 4 follows the 41-category CUAD taxonomy with citations better than GPT-4o in 2026 benchmarks. Legal-BERT fine-tuned is faster at scale but requires 1000+ labeled contracts per category. Start with Sonnet 4 few-shot; graduate to Legal-BERT at 50k+ contract volume.

Alternatives: GPT-4o, Gemini 2.5 Pro, Legal-BERT fine-tuned

Sonnet 4 with structured output mode reliably produces typed JSON (amounts, dates, durations). GPT-4o is nearly as good and has native Structured Outputs. Test both on your specific clause types - the difference is often within eval noise.

Alternatives: GPT-4o, Gemini 2.5 Pro

Relational queries dominate contract Q&A (show me all contracts where indemnity cap > $10M). Postgres handles this plus vector search for similarity. Snowflake if you already have a data warehouse and want legal data there.

Alternatives: Snowflake, BigQuery, Pinecone + Postgres

CUAD (41 clause types across 510 contracts) is the standard benchmark. LEDGAR has more categories but less community traction. Start with CUAD, add 20-30 company-specific clauses (e.g. 'MFN pricing', 'data residency') over time.

Alternatives: LEDGAR, Custom, Atticus

Paralegals need to see the PDF with highlighted extraction spans side-by-side with structured output. Generic tools (Airtable, Notion) lose the PDF context and slow review down 3-5x.

Alternatives: Airtable, Ironclad review, Notion

Whole-contract LLM vs clause-by-clause

Feeding the entire contract to the LLM for extraction is simpler but loses accuracy on long MSAs (30+ pages). Splitting into clauses first and extracting per-clause gives 8-15% higher F1 at the cost of 3-5x more LLM calls. For high-value contract types (MSA, data processing agreement), always split. For short NDAs, whole-contract is fine.

Structured output vs freeform with post-parsing

Anthropic Tool Use and OpenAI Structured Outputs guarantee JSON schema compliance but slightly degrade clause-attribute recall on edge cases. Freeform output + regex post-parse catches more edge cases but fails 2-5% of calls on invalid JSON. Use structured mode for amounts, dates, durations; freeform for narrative fields.

Citation at span-level vs clause-level

Span-level citations (exact characters) are what lawyers want but add 20-30% parsing cost and sometimes break on tables. Clause-level citations (clause ID + page number) are easier to implement and sufficient for most workflows. Upgrade to span-level for high-stakes reviews (M&A due diligence, regulatory filings).

Extraction says 'unlimited indemnity' but the clause actually has a carve-out capping it

Mitigation: Always extract the full clause text alongside the attribute. Require legal review on any attribute marked 'unlimited', 'perpetual', 'exclusive', or 'irrevocable'. These are the clauses that cost companies money when extraction is wrong.

Model cannot find a citation span, so hallucinates the clause

Mitigation: Require exact-span citations as a contract of the extraction. If the model cannot locate the claimed fact in the source, reject the extraction and route to review. Never store an attribute without a verified citation.

Amendments and redlines change the effective terms but pipeline processes only the base contract

Mitigation: Detect amendments (explicit reference to a parent agreement, 'this amends', redline marks). Process in chronological order and merge. Flag contracts where amendment-chain reconciliation failed for human review.

Non-English contracts processed with English-tuned prompts yield garbage

Mitigation: Detect contract language first. Route Spanish, French, German, Mandarin, and Japanese contracts through language-specific prompts. Use Gemini 2.5 Pro (strongest multilingual) or fine-tuned local models. Do not silently run English prompts on Japanese contracts.

Attorney-client privilege leaked to third-party model provider

Mitigation: Use zero-data-retention endpoints (Anthropic ZDR, OpenAI ZDR, Vertex AI). Sign BAA or DPA. Redact client names, matter numbers, and privileged communications before sending. For highly sensitive contracts (M&A, litigation), use self-hosted models (Llama 3, Mistral Large).