Coze Loop Documentation

The Coze Loop backend is organized into 6 Domain-Driven Design (DDD) business modules located in backend/modules/. Each module has a specific responsibility:

• data: Data management, including dataset CRUD and data import/export.
• evaluation: Evaluation management, including evaluation sets, evaluators, and experiment execution.
• foundation: Core capabilities such as users, permissions, and general configurations.
• llm: LLM integration, including model calls and Eino framework integration.
• observability: Observability features like Trace collection, querying, and analysis.
• prompt: Prompt management, including Prompt CRUD, versioning, and Playground.

CozeLoop is a professional AI Prompt Engineering and Evaluation platform designed for developers and researchers. It provides a complete toolchain for AI application development, including prompt development, model evaluation, experiment management, and observability monitoring.

Core Features:

• Prompt Engineering: Visual prompt editor with templates/variables, a Playground for real-time testing, and version control.
• Evaluation System: Management of test datasets, custom evaluators (metrics and rules), and batch experiment execution.
• Observability: Request chain tracing, real-time performance analysis, and structured log management.
• Tag Management: Resource classification, tag-based access control, and batch operations.

Coze Loop is a platform-level solution for AI Agent development and operations (DevOps). It provides lifecycle management capabilities including:

• Prompt Development: Writing, debugging, optimizing, and version management via a visual Playground for real-time interaction and model comparison.
• Evaluation: Systematic, multi-dimensional automated testing of Prompts and Agents (e.g., accuracy, conciseness, and compliance) using managed evaluation sets and experiments.
• Observability: Full-link visualization of execution processes, capturing Prompt parsing, model calls, tool execution, and intermediate results/exceptions via SDK Trace reporting.
• Model Support: Integration with providers like OpenAI and Volcano Ark.

The backend is a Go service located in backend/ with the module path github.com/coze-dev/coze-loop/backend. It follows Domain-Driven Design (DDD) principles. Each business module under modules/<domain>/ is structured as follows:

• application/: Application services for use-case orchestration and Dependency Injection (DI) using Wire.
• domain/: Domain models including entities, repository interfaces, and services. The domain layer must not reference the infra layer.
• infra/: Infrastructure implementations such as repository implementations, RPC, MQ, and storage.
• pkg/: Module-specific utilities.
• consts/: Module constants.

Dependency Direction: api/ $\rightarrow$ application/ $\rightarrow$ domain/ $\leftarrow$ infra/.

The frontend architecture follows a strict 6-level dependency hierarchy to maintain modularity. A critical rule is that higher-level packages can only depend on lower-level packages; reverse dependencies are prohibited.

Dependency Hierarchy

The @cozeloop/prompt-pages package provides a full workflow for prompt lifecycle management:

• Prompt List Management: Browse the prompt library, use advanced search (filter by creator, name, tags), perform batch operations (create, edit, copy, delete), and access execution traces.
• Prompt Development: An interactive editor with version management, multi-modal support (image input), real-time testing, and integrated execution history.
• Prompt Playground: An experimentation environment for rapid testing without saving to the library, supporting templates and model selection.
• Observability: Integrated execution traces, performance metrics (latency, tokens), and debug information via the observation system.