Architecture
Pipeline Overview
- Prompts are created in workspace context
- Queue jobs are scheduled per model provider
- Agent workers execute prompts through provider interfaces
- Responses and citations are extracted and normalized
- Analytics are computed and materialized for dashboard queries
Design Principles
- Provider isolation in workers
- Deterministic prompt run metadata
- Source-linked analysis records
- Separate operational and analytical storage
Data Separation
- Postgres: workspace, user, and scheduling state
- ClickHouse: response/citation analytics and metrics
- Redis: queue orchestration and transient job state