Dike logoDocs

1. Security & Data Privacy

Does Dike train models on customer queries or uploaded contracts?

No. Dike enforces a strict zero-data-retention policy for model training. Customer prompts, search queries, and contract excerpts transmitted through /v1/*, /mcp, or gRPC are never used to train foundation models.

Where is customer data processed?

API requests are processed in secure cloud regions with end-to-end TLS 1.3 encryption in transit and AES-256 encryption at rest in our primary data stores.


2. Arabic Language & Regional Dialects

How does Dike handle colloquial Arabic or spelling variants?

Dike's retrieval engine utilizes state-of-the-art multilingual legal embeddings fine-tuned on diverse Arabic texts. Queries undergo Unicode NFKC normalization, Alef/Hamza folding, and Arabic-Indic digit unification. While statutory terms should ideally be formulated in Modern Standard Arabic (MSA / الفصحى), conceptual queries with regional terminology are mapped to statutory synonyms via dense semantic embeddings.

Can I ask questions in English?

Yes. All endpoints (/v1/search, /v1/reason) support cross-lingual queries. You can submit an English query (e.g. "What is the statutory limitation period for commercial debt in Egypt?"), and Dike will retrieve Arabic statutes and generate an English or Arabic answer based on your "language" parameter.


3. Grounding & Hallucinations

What makes Dike different from standard RAG?

Standard Retrieval-Augmented Generation (RAG) feeds retrieved chunks into an LLM and trusts the model's output. Generalist LLMs frequently hallucinate article numbers or cite overruled laws even when provided with context.

Dike adds an active Grounding Verification Layer that intercepts reference tokens ([[REF:canonical_id#article]]) in real time, validates them against the candidate chunk set, and strips any reference that cannot be deterministically verified against indexed law.

Why do some answers return grounding.used: 0?

If the indexed corpus lacks authoritative statutory or case law evidence to support a decisive answer, the hallucination guard prevents the model from guessing. The response includes grounding.used: 0 and the X-Dike-Grounding: weak header so your application can render a disclaimer.


4. Enterprise & Custom Deployments

Yes. Enterprise plans support dedicated private indexes for searching internal corporate bylaws, standard contracts, and historical arbitration awards alongside the public MENA legal corpus.

Is Dike available for on-premise or private VPC deployment?

Enterprise customers can deploy Dike as containerized microservices within dedicated AWS, Azure, or GCP Virtual Private Clouds. Contact [email protected] for architecture specifications.

On this page