All articles
Case studyBolder Team

Arabic legal research with citations

Judger is an Arabic-first legal chatbot that lets Jordanian lawyers ask questions in plain Arabic and get cited answers from case law and statutes. It cut research time on standard case law queries by about 65%.

Arabic legal research with citations

Jordanian law firms work with a large volume of case law, statutes and legal commentary, most of it in Arabic. Junior associates spend hours searching for the right precedent. Senior partners then spend hours checking that work.

We built Judger for a Jordanian legal services client. Lawyers ask questions in natural Arabic and get answers drawn from a curated corpus of Jordanian and regional Arab law, with the source cited every time.

The problem

  • Legal research in Arabic is slow, and the material is spread across several sources.
  • Existing tools are either built for English first or don't know the legal domain well.
  • Associates needed to find relevant rulings without hours of manual search.
  • Partners needed to be able to trace every AI answer back to a real source.

What we built

Arabic-first retrieval pipeline

We built the retrieval pipeline around Arabic legal text from the start:

  • Ingestion. Structured parsing of Jordanian statutes, court rulings and regulatory texts.
  • Chunking. Splits follow Arabic sentence structure and legal clause boundaries instead of fixed token counts.
  • Hybrid retrieval. Dense and sparse search combined, with re-ranking tuned on real legal query patterns.
  • Citations. Every answer names the exact source document and clause.

Conversational interface

Lawyers can have a multi-turn conversation. They ask a follow-up, narrow the scope or ask for a comparison, without starting over.

Judger understands both Jordanian legal Arabic and formal Modern Standard Arabic. When a question is ambiguous, it asks for clarification before it answers.

Evaluation

We evaluated the system continuously during development with BEval Studio:

  • Automated scoring for answer relevance, citation accuracy and hallucination rate
  • Practicing Jordanian lawyers reviewing a sample of outputs by hand

The final system reached over 95% citation accuracy on held-out test queries.

Results

MetricResult
Research time, standard case law queriesDown about 65%
AdoptionAssociates use it daily as their first-pass research tool
Hallucinated citationsZero in post-launch monitoring over the first 60 days

In legal work, an answer without a source isn't much use. Every Judger answer points to the document and clause it came from, so a partner can check it.

The client's name is withheld under NDA. Judger is in production in Jordan.

Building something similar?

Tell us what you're working on. We'll tell you how we'd approach it.

Partner with us →