Skip to content

Knowledge Platform — Deployment

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Runtime model

Concern Choice
Deployment unit KnowledgeHost, one image from the microservice template; six modules
Hosting AKS, 2 to 10 replicas; ingestion workers scale separately inside the host by partition
Data knowledge-sql (metadata, facts, snapshots, artifacts metadata, memory), vector-index (Qdrant reference profile; Azure AI Search, PgVector or SQL Server per ADR-0013), artifact-blob (content by hash), cache
Messaging MessagingModel outbox and inbox; Orleans profile by default
Identity workload identity; source and store credentials brokered
Ingress gRPC from runtime workers and the tool executor; GW-03 for Studio reads
Egress sources through the Integration platform, model gateway for embeddings, Governance; everything else denied

Topology by profile

Profile Knowledge host placement Stores
saas-default ConnectSoft AKS in the tenant's region tenant-isolated schemas and collections in regional stores
hybrid inside the customer's network next to repositories and documents; outbound-only to the Control plane customer-hosted stores
on-prem customer cluster customer stores; Qdrant or PgVector typical
hybrid-compute as saas-default as saas-default; embeddings may route to tenant-private pools 
flowchart LR
    subgraph Region["Tenant region"]
        KH["KnowledgeHost"]
        SQL["('knowledge-sql')"]
        VEC["('vector-index')"]
        BLOB["('artifact-blob')"]
    end
    KH --> SQL & VEC & BLOB
    KH -- "Embed, Rerank" --> MG["ModelGatewayHost"]
    KH -- "connections" --> IH["IntegrationHost"]
    RT["AgentRuntimeWorker"] --> KH
Hold "Alt" / "Option" to enable pan & zoom

Scaling

Signal Action
retrieval latency above target add replica; cache hot collections
ingestion backlog (snapshots in Taken) scale ingestion partitions; embeddings are rate-limited by the model gateway
vector index size collection sharding per tenant and collection in the vector profile
artifact bytes blob tiering (cold after retention threshold)

Configuration and secrets

Options sections Knowledge:Host, Knowledge:VectorProfile, Knowledge:Ingestion, Knowledge:Retrieval, Knowledge:Graph, Knowledge:Context, Knowledge:Artifacts, Knowledge:Memory, Knowledge:Repositories. Secrets for the three stores and the transport in Key Vault through workload identity; source credentials never configured, always brokered.

Health and rollout

/knowledge/health with per-store and connector health; rolling updates drain ingestion jobs at stage boundaries; an embedding profile version change triggers index rebuilds that run while the old revisions serve.

NFR class, residency and recovery

pilot-interactive for retrieval and context, pilot-batch for ingestion and memory (D30). Stores live in the tenant's region (D31). Restore order: knowledge-sql after the control and runtime stores, then artifact-blob, then the vector index, which is rebuilt from published snapshots when its backup is older than the metadata.