ADR-007: AI — LLM Choice
Status: Accepted Date: May 2026
Context
BayanCore requires LLMs for AI assistants, document processing, and Arabic NLU — all hosted in KSA.
Decision
Multi-model strategy:
- Primary: Open-source models (Llama, Mistral) hosted on OCI GPU instances
- Arabic NLU: Fine-tuned models for Saudi business dialect
- Embeddings: Text embedding models for RAG pipeline
Hosting: All inference runs in OCI Riyadh — no external API calls for production data.
Consequences
- Positive: Full data residency, cost control at scale, no vendor lock-in
- Trade-offs: Requires GPU infrastructure management
- Risks: Model quality may lag behind proprietary models
Alternatives Considered
- OpenAI API: Best quality but data leaves KSA
- Azure OpenAI: KSA region not available at time of decision
- Local LLM Only: Limited Arabic support without fine-tuning