إنتقل إلى المحتوى الرئيسي

ADR-008: Vector Store — OCI Search with OpenSearch

Status: Proposed Date: 2026-05-17

Context​

BayanCore's AI features (Ask, Act, Automate tiers) require a vector database for the RAG (Retrieval-Augmented Generation) pipeline. The system must store embeddings of company documents, policies, and business data to enable semantic search and context-aware AI responses. All vector data must remain in KSA to comply with PDPL data residency requirements.

Since we selected MariaDB as the primary ERP database (ADR-003), we cannot use pgvector. We need a dedicated vector search solution that is fully managed, KSA-compliant, and integrates well with our OCI-hosted AI inference models.

Decision​

OCI Search with OpenSearch is selected as the vector search engine for BayanCore's RAG pipeline.

Rationale:

  • Fully managed service by OCI — minimal operational overhead
  • Available in OCI Riyadh region — full PDPL compliance
  • Native vector search with k-NN plugin for billion-scale similarity search
  • Hybrid search: combines vector similarity with full-text search for better results
  • Built-in RAG pipeline support with OCI Generative AI Agents integration
  • LangChain integration for simplified AI application development
  • Supports Cohere and Llama embedding models (both OCI-hosted)
  • Enterprise-grade security with encryption, access controls, and auditing

Deployment Configuration:

  • Single OpenSearch cluster in OCI Riyadh
  • Vector indexes for company document embeddings
  • Full-text indexes for keyword search fallback
  • Integration with OCI Generative AI for embedding generation (Cohere embed-multilingual-v3)
  • Data ingestion pipeline: documents → chunking → embedding → OpenSearch index

Consequences​

  • Positive: Fully managed, KSA-compliant, RAG-ready, hybrid search capabilities, less ops overhead than self-hosted alternatives
  • Trade-offs: Tied to OCI ecosystem, OpenSearch-specific query syntax
  • Risks: Service availability depends on OCI region health, embedding model costs

Alternatives Considered​

  • Qdrant on OKE: More flexible vector features but requires managing Kubernetes deployment and increases operational complexity
  • Oracle AI Vector Search (Oracle DB 26ai): Powerful but requires Oracle Database license (expensive) and conflicts with MariaDB decision
  • pgvector on PostgreSQL: Would require running PostgreSQL alongside MariaDB, adding unnecessary complexity
  • Milvus self-hosted: Apache-licensed and scalable but requires full self-management on OCI compute