---
title: "Announcing Dike: The Legal Source Layer for AI"
date: "2026-08-11"
description: "Why we built Dike to provide structured, verifiable legal data directly to AI agents."
author: "Dike Team"
readTime: "5 min read"
heroImage: "/images/artwork/universal-source.png"
tags: ["Announcement", "Legal Tech", "AI Infrastructure"]
---

Today we're introducing Dike, a structured API that gives AI agents authentic access to primary legal sources across the Middle East and North Africa.

We started this because we kept running into the same problem from different directions: an agent asked about a Saudi labor dispute, a compliance tool checking Egyptian arbitration deadlines, a contract review workflow trying to compare gratuity rules across two jurisdictions. In every case, the model had an opinion. What it didn't have was the statute.

> The model always has an answer. The question is whether it has a source.

## The problem with legal AI right now

Language models are genuinely good at sounding like they know the law. That's exactly what makes them risky to use for it. Ask a model for a specific article number, a filing deadline, or the current text of a regulation, and it will answer fluently whether or not it actually has the right information in its weights. A few ways this shows up in practice:

Models invent statutory numbers and citations that simply don't exist, dressed up with plausible-sounding sub-provisions. They quote articles that were repealed or amended years ago, because their training data is a snapshot and the law isn't static. And even when an answer happens to be correct, there's usually no way to check it, no citation anchor, no link back to an official source, just a confident paragraph of text.

For English-language, US and EU law, there are now a handful of legal APIs and datasets trying to close this gap. For Arabic-language MENA law, there wasn't really anything comparable. Global legal data providers treat the region as an afterthought at best, often relying on machine translation that strips out the citation structure a lawyer actually needs. Web search is worse: blog posts, law firm marketing pages, and outdated summaries, with no way to tell which ones are still accurate.

## What Dike actually does

Dike turns statutes, administrative regulations, court rulings, and official gazette entries into a single, queryable, bilingual API. Instead of an LLM guessing from compressed parameters, an agent can call Dike's endpoints, `search`, `cite`, `graph`, and `documents`, and get back real Arabic legal text with English alongside it, tied to a official source ID.

A few things we built specifically because the naive approach doesn't hold up in this domain:

**Hybrid search that handles Arabic legal terminology.** Pure keyword search misses paraphrased queries; pure vector search misses exact article references. We run both together, because legal queries genuinely need exact-match precision on citations and semantic matching on everything else.

**Citation graphs, not flat documents.** A statute doesn't exist in isolation. It has amendments, it has implementing regulations, and it gets interpreted by court rulings that cite it. We index those relationships so an agent can walk from a law to the regulation that implements it, to a court ruling that applied it, and back.

**Verification against the live index, not a cache.** Every citation an agent generates through Dike gets checked against currently indexed text, so a repealed article gets flagged as repealed instead of quoted as if it's still good law.

## Where we're starting

We're not trying to index the whole world on day one, and we don't think anyone honestly can. We're starting deep rather than wide: Saudi Arabia, the UAE, and Egypt, with GCC and the rest of MENA next.

> A legal AI product needs the full depth of a jurisdiction, not a thin layer across a hundred of them.

Every amendment, every implementing regulation, every relevant ruling in three countries beats a shallow pass across a hundred that's too thin to be trustworthy anywhere.

If you're building an AI agent that touches Arabic legal text in any way, from contract review to compliance monitoring to a general-purpose research assistant, we'd like to hear what you're running into. The API is live, and we're actively expanding coverage based on what people building on it actually need.