Architecture
============

- Control Plane: router, planner, authority model, buses (NATS/Redpanda)
- Model Plane: vLLM OpenAI server, Gorilla bridge
- Memory Plane: vector, graph, GraphRAG orchestrator
- Tools Plane: MCP registry, dynamic tool creation (gorilla/http)
- IO Plane: gateway proxy (w/ RAG), Kong external ingress
- Observability: Prometheus metrics per service; OTel traces to console

Key Ports
- Gateway: 8088
- Router: 7000
- Tools MCP: 7001
- Vector: 7010, Graph: 7011, GraphRAG: 7012, Embedders: 7013, Ranker: 7014, Summarizer: 7015
- Gorilla bridge: 8089

Data Flow (RAG)
1) /v1/rag/completions → gateway embeds query_text (7013) → GraphRAG/query (7012)
2) GraphRAG searches vectors (7010), expands neighbors (7011), optional rerank (7014)
3) Gateway caches result in Dragonfly/Redis and prepends [CONTEXT] → upstream completions

Buses
- Set NATS_URL/KAFKA_BOOTSTRAP to emit ops events on route decisions.