AI & Automation · Meridian Legal

A RAG assistant lawyers actually trust, with cited answers

Meridian's lawyers spent hours digging through thousands of contracts. We built a retrieval system that answers questions with clause-level citations, taking research from hours to seconds.

Client

Meridian Legal

Industry

Legal services

Timeline

10 weeks

Year

2026

0%

accuracy on evaluation set

0K

contracts indexed and searchable

0%

less time spent on research

01 · The challenge

Meridian Legal manages over 40,000 contracts for its clients. Finding precedent language, unusual clauses, or conflicting terms meant associates manually searching PDFs, sometimes for a full day per request.

Accuracy requirements were brutal. A wrong answer in legal work is worse than no answer, so a generic AI chat tool was never an option. Every answer needed a source the lawyer could click and verify.

02 · The solution

We built a clause-level RAG pipeline. Instead of splitting documents into arbitrary chunks, contracts are parsed into their legal structure: parties, definitions, obligations, termination clauses. Retrieval happens at the clause level, and every answer links to the exact clause in the original document.

A reranking model orders results by legal relevance, not just similarity, and an evaluation set of 300 real research questions keeps quality measurable. Accuracy on that set went from 71% with naive chunking to 94% with the final pipeline.

03 · How we did it

The approach

1

Parse the legal structure

Custom parsing broke 40,000 contracts into typed clauses, preserving cross-references and definitions instead of arbitrary chunks.

2

Build retrieval for law

Hybrid search combining embeddings and keyword matching, with a reranker tuned on lawyer-labeled relevance judgments.

3

Cite everything

Every answer includes clause-level citations that open the source document at the exact paragraph, so verification takes one click.

4

Measure with real questions

A 300-question evaluation set built with the partners, scored on every release, so accuracy is a number and never a feeling.

Research that took an associate a full day now takes seconds, and every answer shows me exactly which clause it came from. That citation is why we trust it.

Sarah Mitchell

COO, Meridian Legal

The stack

Anthropic ClaudeLlamaIndexWeaviateNext.jsPythonFastAPIPostgreSQLAzure

Want results like this?

Tell us where your numbers are today and we will show you what they could look like.