Legal Contract Intelligence

A legal team wants natural-language search over ten years of internal documents and contracts - questions like which of our master agreements cap liability below twelve months of fees. The hard parts are not the LLM. They are getting clean text out of scanned PDFs, retrieving at clause granularity rather than document granularity, knowing which version of an agreement was in force on a given date, and refusing to answer when the corpus does not support one. Permission filtering matters here too, but it is covered end to end on the Enterprise Knowledge Agent problem - this one is about the document layer.

“Every answer cites the clause, not the document. A contract answer without a pinpoint citation is unusable to a lawyer, who has to verify it before relying on it. The retrieval unit is the clause; the document is only the container.”

• 1OCR is the accuracy ceiling - Everything downstream inherits OCR error. Budget for a layout-aware model, keep per-page confidence, and route pages below threshold to a review queue rather than indexing garbage. Faxed amendments, stamped exhibits and signature pages are where accuracy collapses.
• 2Clause-level chunking, not fixed windows - Split on the contract's own structure - numbered clauses, schedules, exhibits - and carry the clause path in metadata. A 512-token sliding window cuts an indemnity in half, and the retrieved fragment can then mean the opposite of the clause.
• 3Defined terms travel with the chunk - Contracts define their own vocabulary. Resolve defined terms at index time and attach the definitions a clause depends on, or the model reads the clause with the everyday meaning of the word rather than the contractual one.
• 4Hybrid retrieval, weighted toward exact terms - Legal queries carry exact strings: party names, clause numbers, Section 11.3, Force Majeure. Dense vectors alone miss them. Merge BM25 and vector with Reciprocal Rank Fusion, then cross-encoder rerank the top 100 clauses.
• 5Version and effective-date resolution - An amendment silently replaces a clause. Store the agreement graph and resolve questions against an as-of date, defaulting to today and stating that date in the answer. Without this the system confidently quotes superseded text.
• 6Mandatory pinpoint citation and abstention - Every sentence carries document, clause and page, with a verbatim quote of the operative sentence. Below the score threshold, or where two versions conflict, the system lists candidates instead of picking one.

• 1Implement OCR review queue - Flag low-confidence OCR pages for manual review rather than indexing them. This prevents silent data loss and maintains high accuracy in the indexed corpus.
• 2Use clause-level chunking - Index clauses based on the document's own structure to ensure accurate retrieval. Avoid fixed token windows that can split clauses and change their meaning.
• 3Resolve defined terms at index - Attach definitions to clauses during indexing to preserve contractual meanings. This prevents the model from misinterpreting terms with everyday meanings.
• 4Employ hybrid retrieval strategy - Combine exact term matching with semantic retrieval to handle legal queries effectively. This approach captures both precise legal terms and broader semantic context.

• 1Silent OCR loss - A page that failed to OCR is simply absent from the index, so the system answers confidently from the pages that worked. Mitigated by reconciling page counts per document and alerting on any document with unindexed pages.
• 2Version confusion - Amendments and side letters change operative terms without changing the original file. Mitigated by the agreement graph, an explicit as-of date on every answer, and surfacing the amendment chain in the citation.
• 3Confident answers from a thin corpus - Absence of a clause is not absence of the obligation - it may live in a document that was never scanned. Mitigated by an abstain threshold and by reporting corpus coverage per counterparty alongside the answer.

Google Stack: Document AI (OCR + layout parsing) → Cloud Storage (page images + extracted text) → Vertex AI embeddings + Vertex AI Vector Search (clause index) → BigQuery (agreement graph + audit) → Gemini (synthesis) → Cloud DLP (PII detection on export)

Azure Stack: Azure AI Document Intelligence (OCR + layout) → Blob Storage → Azure AI Search (hybrid BM25 + vector with the semantic reranker) → Azure SQL (agreement graph) → Azure OpenAI GPT-4o (synthesis) → Microsoft Purview (retention + audit)

AWS Stack: Amazon Textract (OCR + tables) → S3 → Amazon Bedrock Knowledge Bases with OpenSearch Serverless (hybrid retrieval) → DynamoDB (agreement graph) → Bedrock Claude (synthesis) → CloudTrail (retrieval audit)

Other Options: Docling or Unstructured.io for parsing + Qdrant or Elasticsearch for hybrid retrieval + a self-hosted Llama for firms that cannot send text to a cloud | Buy rather than build - Kira, Luminance or Harvey - and be ready to argue the build-versus-buy line, because for a single legal team it is often the right call.