coze-studio

Provides a React 18-based visual canvas IDE for composing conversational AI agents by connecting LLM models, RAG knowledge bases, plugins, and workflow nodes without code. Uses a FlowGram engine to render and manage directed acyclic graphs of agent logic, with Zustand state management for real-time canvas synchronization and a Thrift IDL layer enforcing strict type contracts between frontend and Go backend services that execute the composed workflows.

Abstracts LLM provider APIs (OpenAI, Volcengine, and others) through a unified model service layer that manages model lists, credentials, and request routing. The backend uses Go's Hertz HTTP framework with domain-driven service handlers that normalize provider-specific request/response formats into a common interface, allowing agents to switch models or providers without workflow changes.

Exposes agent functionality through OpenAPI endpoints for chat session management and a Chat SDK (TypeScript/Python) for application integration. The OpenAPI spec is auto-generated from Thrift IDL, providing standard REST endpoints for creating sessions, sending messages, and retrieving traces. The Chat SDK wraps these endpoints with convenience methods, error handling, and streaming support for real-time agent responses.

Provides Docker Compose configurations for local development and Kubernetes Helm charts for production deployment. The Docker Compose setup includes all services (frontend, backend, MySQL, Redis, Elasticsearch, vector databases) with environment variable configuration. Helm charts abstract Kubernetes resources (Deployments, Services, ConfigMaps, Secrets) and enable parameterized multi-environment deployments (staging, production) with different resource limits and replica counts.

Provides a resource management system for uploading, indexing, and retrieving documents through a RAG pipeline built on the Eino framework. Documents are embedded using configurable vector models, stored in vector databases (Milvus, OceanBase, or similar), and retrieved via semantic search with BM25 hybrid ranking. The backend Go services handle chunking, embedding, and retrieval orchestration, while the frontend provides UI for knowledge base CRUD and search testing.

Enables agents to invoke external tools and APIs through a plugin registry system where each plugin defines a Thrift-based schema specifying inputs, outputs, and execution logic. The backend maintains a plugin service that validates requests against schemas, handles authentication/credentials, and orchestrates execution via HTTP or gRPC. Plugins can be built as standalone services or embedded Go modules, and the frontend provides UI for plugin discovery, configuration, and testing.

Manages the complete agent lifecycle from creation through deployment, including version control, publishing to registries, and deployment to production environments. The backend stores agent definitions (prompts, workflows, RAG bindings, plugins) in MySQL, tracks version history, and provides APIs for publishing agents as immutable releases. The frontend IDE includes publish workflows, deployment configuration UI, and agent marketplace browsing for discovering and importing published agents.

Manages chat sessions between users and deployed agents, capturing full execution traces including LLM calls, tool invocations, RAG retrievals, and workflow steps. Sessions are stored in MySQL with Redis caching for active sessions, and the backend exposes OpenAPI endpoints for session creation, message sending, and trace retrieval. The frontend provides a chat UI with side-by-side execution trace visualization, allowing developers to inspect intermediate states and debug agent behavior.

Provides a visual canvas for designing complex agent workflows with nodes for LLM calls, tool invocations, RAG retrievals, and control flow (conditionals, loops). The FlowGram engine renders the workflow as a directed acyclic graph (DAG) with drag-and-drop node composition, and the backend interprets the workflow definition to execute steps sequentially or in parallel. Workflows support variable binding, expression evaluation, and error handling with fallback paths.

Provides a unified development environment using Rush.js for TypeScript/React frontend package management and Docker Compose for local backend/infrastructure deployment. The monorepo separates frontend IDEs (Agent IDE, Studio, Project IDE), backend services (API handlers, domain services), and shared infrastructure (MySQL, Redis, Elasticsearch, vector databases). Developers can run the entire stack locally with a single `make web` command, with hot-reload for frontend changes and automatic backend service restart on code changes.

Defines all frontend-backend communication through Thrift IDL files, ensuring strict type safety and contract enforcement across the monorepo. The Thrift compiler generates TypeScript types for the frontend and Go structs for the backend, preventing runtime type mismatches and enabling automatic API documentation. Changes to IDL files trigger validation in CI/CD pipelines, and breaking changes are flagged before merge.

Organizes backend services using Domain-Driven Design (DDD) principles with clear separation between API handlers, domain services, and infrastructure adapters. Each domain (agents, workflows, RAG, plugins) has its own service layer that encapsulates business logic, with infrastructure adapters for databases, caches, and external APIs. The Hertz HTTP framework routes requests to handlers that delegate to domain services, ensuring business logic is decoupled from HTTP concerns.