ClauseIQ — Legal Contract Intelligence Assistant

A commercial law firm handling 200+ contracts per month had junior associates spending 3–4 hours on initial review of each document — reading for non-standard terms, liability exposure, and missing clauses. ClauseIQ automated that first pass: associates upload a contract, and within 90 seconds receive a structured risk report, redline suggestions, and a chat interface where they can ask 'What's the termination notice period?' and get a cited answer.

• 1

  Legal documents use precise, domain-specific language where a single word change ('reasonable' vs. 'sole') can shift million-dollar liability — general-purpose LLMs hallucinated too often.

• 2

  Contracts reference external documents, exhibits, and defined terms recursively — the system needed to understand cross-references, not just paragraphs in isolation.

• 3

  Client confidentiality rules meant no data could leave a private infrastructure boundary — cloud LLM APIs were not an option.

• 4

  Lawyers needed to trust the output — every claim had to cite the exact clause location, not just the summary.

We deployed a private RAG (retrieval-augmented generation) stack running on the firm's own servers. Contract documents are chunked at the clause level, embedded with a fine-tuned legal-domain model, and stored in a private Qdrant vector database. Queries retrieve the top-k most relevant clauses, which are passed to a locally hosted Llama 3 model fine-tuned on contract review data. The risk detection layer uses a separate classifier trained on 12,000 annotated clauses to flag 23 risk categories (uncapped liability, automatic renewal, unilateral amendment rights, etc.) with confidence scores. All citations include page number, section reference, and the verbatim sentence.