How lmnr-ai/lmnr works
Project type and purpose
• Laminar is a self-hostable, open-source full-stack platform for tracing and evaluating AI applications.
• Key features from the README:
• OpenTelemetry-based automatic tracing via gRPC
• Parallel evals SDK
• Data export to datasets
• Dashboards for traces, metrics, labels
• Core backend in Rust, agent manager in Python, frontend in Next.js
File structure
• Top-level services:
• agent-manager (Python gRPC service & browser agent)
• app-server (Rust core backend, gRPC endpoints, RabbitMQ consumers/producers, Postgres/ClickHouse drivers)
• frontend (Next.js UI + API routes)
• docker-compose files orchestrating Postgres, ClickHouse, RabbitMQ, agent-manager, app-server, frontend
• Recognize patterns:
• Microservices (agent-manager vs. app-server vs. frontend)
• Event-driven, queue-based trace ingestion (RabbitMQ)
• gRPC for high-throughput telemetry ingestion
• REST/Next.js API for UI interactions
• OLAP (ClickHouse) vs. OLTP (Postgres)
README for architecture insights
• SDKs/instruments → gRPC → app-server
• RabbitMQ for decoupled trace processing
• Postgres for application state
• ClickHouse for analytics
• Next.js frontend consuming REST API
• Client layer: TS & Python SDKs + instrumented AI frameworks
• agent-manager service (Python gRPC)
• app-server service (Rust gRPC + RabbitMQ producer/consumer + REST API)
• Message broker: RabbitMQ
• Databases: Postgres (OLTP), ClickHouse (OLAP)
• Next.js frontend (UI + API routes)
• External services: AI providers (OpenAI, Anthropic, LangChain, Scrapybara)
• SDK → gRPC → app-server / agent-manager
• agent-manager → RabbitMQ queue → app-server consumer
• app-server → Postgres for metadata & control data
• app-server → write trace batches to ClickHouse
• app-server → consume from RabbitMQ for label/eval events
• Next.js frontend → REST API → app-server
• Frontend static assets served via Docker container
• External AI services called by SDKs/agent-manager
c. Architectural patterns & principles:
• Microservices separation of concerns
• Event-driven / asynchronous processing via message queue
• Use of gRPC for low-latency instrumentation
• CQRS-style separation: Postgres for writes/state, ClickHouse for read/analytics
• API layering: gRPC for ingestion, REST for UI
d. Key technologies & frameworks:
• Rust (app-server)
• Python (agent-manager)
• Next.js / React (frontend)
• gRPC + protobuf
• RabbitMQ
• Postgres
• ClickHouse
• Docker & Docker Compose
Detail without overcomplicating
• Distinct service and its role.
• Data stores separately (Postgres vs. ClickHouse).
• External dependencies (AI providers, browser agent manager).