Aegis-RAG — System Architecture

Aegis-RAG — System Architecture An architecture diagram generated by Archify. Web Console · React SPA · nginx :8080 · Architecture component Web Console React SPA · nginx :8080 Aegis-RAG API · FastAPI · hexagonal :8000 · Local stack · docker compose (aegis-net) — air-gapped by design › Ingress: auth · rate limit · headers · OWASP LLM Top 10 Aegis-RAG API FastAPI · hexagonal :8000 OWASP LLM Top 10 Redis · rate-limit store :6379 · Local stack · docker compose (aegis-net) — air-gapped by design Redis rate-limit store :6379 ChromaDB · vector store :8000 · Local stack · docker compose (aegis-net) — air-gapped by design ChromaDB vector store :8000 Ollama · LLM · llama3.2 :11434 · Local stack · docker compose (aegis-net) — air-gapped by design Ollama LLM · llama3.2 :11434 OTel Collector · OTLP · spanmetrics · Local stack · docker compose (aegis-net) — air-gapped by design OTel Collector OTLP · spanmetrics Jaeger · traces :16686 · Local stack · docker compose (aegis-net) — air-gapped by design Jaeger traces :16686 Prometheus · metrics :9090 · Local stack · docker compose (aegis-net) — air-gapped by design Prometheus metrics :9090 Grafana · dashboards :3000 · Local stack · docker compose (aegis-net) — air-gapped by design Grafana dashboards :3000 query · ingest OTLP spans traces scrape span metrics query Local stack · docker compose (aegis-net) — air-gapped by design Ingress: auth · rate limit · headers Legend Frontend Backend Database Message bus External

Security is the architecture

  • • SecurityGateway scores every query before retrieval — prompt injection blocked at 400
  • • OutputSanitizer guards every LLM response (LLM05)
  • • API-key auth (constant-time) + per-key sliding-window rate limit

Hexagonal core

  • • Ports & adapters — swap ChromaDB or Ollama in one wiring file
  • • Query path: security -> retrieve -> generate -> sanitize
  • • Embeddings baked into the image (multilingual MiniLM) — boots fully offline

Observability

  • • OTLP spans -> collector -> Jaeger + spanmetrics
  • • Prometheus + Grafana; exemplars jump from a metric to its trace
  • • Grafana reads traces from Jaeger for one-click drill-down