narender-rk10/openbi

Purpose and Entry Points

OpenBI is a self-hosted AI business-intelligence platform: connect live data sources, define MindsDB-backed agents, query via chat, build dashboards, and distribute reports. The FastAPI process starts at backend/main.py; domain routers are assembled from backend/api/. The React SPA starts at frontend/src/main.tsx and routes from frontend/src/App.tsx. Docker topology is defined in docker-compose.yml.

Application Boundary, Identity, and Persistence

The backend owns tenancy, application state, and authorization; MongoDB is the system of record, accessed asynchronously through backend/core/database.py. JWT authentication, bcrypt password verification, and Fernet encryption for stored connection/LLM credentials are centralized in backend/core/security.py; API dependency enforcement is in backend/api/deps.py. Domain APIs persist projects, connections, agents, chats, dashboards, schedules, exports, and settings in MongoDB, while MindsDB holds executable agent/KB constructs.

Data Connectivity and AI Agent Layer

Connection management translates UI-provided source settings into encrypted MongoDB records and MindsDB database handlers through backend/api/connections.py and backend/services/mindsdb_client.py. Agent creation binds selected tables/connections and optional knowledge bases into MindsDB multi-skill agents via backend/api/agents.py. MindsDB is therefore the boundary to 90+ source handlers, text-to-SQL execution, and RAG retrieval; the OpenBI backend does not directly query customer databases.

Chat Execution Pipeline

frontend/src/hooks/useChat.ts consumes the primary SSE interaction exposed by backend/api/chat.py. Requests can target mentioned agents, use an active agent, or invoke LLM-based routing. The backend restores bounded conversation history from MongoDB, handles greetings locally, streams MindsDB agent output, extracts generated SQL, re-executes it for structured rows, and emits status/SQL/data/answer events. backend/services/llm_client.py abstracts Gemini, OpenAI, and Anthropic for routing and synthesis; optional tracing flows through backend/services/langfuse_client.py.

Dashboard Authoring and Real-Time State

Dashboards store widget SQL, display configuration, cached result data, filters, and versions as application documents behind backend/api/dashboards.py and backend/services/version_service.py. Widgets execute through MindsDB; AI-generated visual/table configurations are produced by backend/services/chart_agent.py and backend/services/table_agent.py. The primary interactive surface is frontend/src/pages/DashboardViewPage.tsx, rendering chart/table/KPI/narrative widgets from frontend/src/components/dashboard/. WebSocket coordination is handled by backend/api/ws.py and backend/services/websocket_manager.py, with Redis pub/sub as the multi-viewer broadcast boundary.

Knowledge Bases and Exports

Knowledge-base APIs in backend/api/knowledge_bases.py ingest uploads or crawled content, prepare supported files, and delegate chunking, embeddings, and vector storage to MindsDB. PDF history/export behavior is represented by backend/services/pdf_service.py; PPTX generation delegates asynchronously to the separate Presenton service through backend/services/pptx_service.py. PDF capture is client-side from dashboard UI, while persisted export/version metadata remains in MongoDB.

Automation and Deployment Services

Schedules are persisted and registered in the in-process APScheduler flow in backend/services/scheduler_service.py; execution refreshes dashboard data and dispatches email, Telegram, or webhook deliveries through backend/services/report_runner.py and backend/services/delivery_service.py. The optional bot is an independent process at telegram_bot/bot.py. Material runtime services are FastAPI/Python 3.11, React/TypeScript, MindsDB, Redis, external MongoDB, optional Presenton, SMTP, Telegram, and Langfuse, containerized by docker-compose.yml.