How would you design a system that receives a user query, uses a performance vector from a module, sends context to an AI component, and returns a global response?

Designing a system that ingests a user query, applies a performance vector, and returns a global response involves several layers. At the front end, an API gateway receives the query and forwards it to a request‑handler service. The handler normalizes the input and calls a performance‑vector module (a microservice that computes a vector representation of the query, perhaps using embeddings). The vector is stored in a fast key‑value store (e.g., Redis) and used to retrieve relevant context from a context‑store service that aggregates knowledge from knowledge bases, logs, or external APIs. The context and vector are then sent to an AI inference engine (e.g., a transformer model hosted on a GPU cluster). The inference engine returns a raw response, which a response‑aggregator service enriches with metadata, applies post‑processing rules, and formats it into a global response. The system uses asynchronous messaging (Kafka) for decoupling, circuit breakers for resilience, and a monitoring stack (Prometheus + Grafana) for observability. Security is handled via OAuth2 and TLS. This architecture scales horizontally, supports A/B testing, and allows independent evolution of each component.