Tác nhân & Đa tác nhân

Opik: Open-Source LLM Observability, Evaluation and Agent Tracing

comet-ml/opik

Opik is an Apache-2.0-licensed platform for tracing, evaluating, debugging and monitoring LLM applications, RAG systems and multi-step agent workflows. It supports self-hosting, a hosted Comet.com option, client SDKs, a REST API, automated evaluations, prompt experimentation and production dashboards.

★ 21,1KSố sao
⑂ 1,7KFork
163Vấn đề đang mở
PythonNgôn ngữ
Apache-2.0Giấy phép
Q@project.QualityScoreĐiểm biên tập

Tổng quan

The repository contains the Opik server backend, web application and components for tracing, datasets, experiments, evaluations, prompt management, online evaluation and agent optimization. It is framework-agnostic and can capture LLM calls, conversations, retrieval steps, tool executions and nested agent activity as trace trees. Evaluation capabilities include code-based metrics, datasets, experiments and LLM-as-a-judge metrics for hallucination detection, moderation and RAG assessment.

Tính năng chính

  • Nested tracing for LLM calls, conversations, retrieval steps, tool executions and multi-step agents
  • Datasets, experiments, code-based metrics and LLM-as-a-judge evaluation
  • RAG evaluation metrics including answer relevance and context precision
  • Hallucination detection and moderation evaluation
  • Prompt management and Prompt Playground
  • Production dashboards and online evaluation rules
  • Feedback scores for traces and spans through the Python SDK or UI
  • PyTest integration for CI/CD evaluation
  • Agent Optimizer and guardrails capabilities
  • Python, TypeScript and Ruby-via-OpenTelemetry client support plus a REST API
  • Hosted and self-hosted deployment options
  • Docker Compose scripts and Kubernetes/Helm deployment guidance

Yêu cầu, cài đặt và bắt đầu nhanh

Python SDK quick start: `pip install opik` or `uv pip install opik`, followed by `opik configure`. For local self-hosting, clone `https://github.com/comet-ml/opik.git`, enter the `opik` directory and run `./opik.sh` on Linux or macOS. On Windows, run `powershell -ExecutionPolicy ByPass -c ".\\opik.ps1"`. The scripts also expose `--infra`, `--backend` and `--guardrails` profiles. Kubernetes installation through Helm is documented for larger-scale deployments. A hosted Comet.com option is also available.

Cách sử dụng

After configuring the Python SDK, import `track` from `opik` and decorate a function with `@track`. Calls to that function, including nested calls, are then logged as traces. For a local installation, the README states that the web interface is available at `http://localhost:5173` after startup. Applications can alternatively log through supported integrations, OpenTelemetry, other client SDKs or the REST API.

Khả năng tương thích của mô hình và trường hợp sử dụng

Opik is described as framework-agnostic rather than tied to one model. The supplied README excerpt explicitly lists integrations involving Anthropic, Amazon Bedrock, BytePlus, Cloudflare Workers AI, Cohere, DeepSeek, Fireworks AI, Google Gemini, Groq and other frameworks or platforms. It also names integrations such as Google ADK, AG2, Agno, Autogen, BeeAI, CrewAI, Dify, DSPy and Flowise AI. OpenAI appears in the repository topics, but explicit OpenAI integration details are not included in the supplied excerpt. MCP support is not stated in the repository metadata.

Ghi chú về giấy phép và rủi ro

Repository metadata identifies the license as Apache-2.0. The README states that the server, web application and core observability and evaluation platform can be self-hosted without a commercial license. Review the repository license before deployment or redistribution: https://api.github.com/licenses/apache-2.0

Phát hành và bảo trì

The README directs users to the changelog and specifically warns that version 1.7.0 includes important updates and breaking changes. The repository metadata reports the default branch as `main` and the latest supplied push timestamp as 2026-08-03T09:50:41Z. A specific current release version is not stated in the repository metadata.

Firecrawl

firecrawl/firecrawl

★ 161,1KTypeScript

LangChain

langchain-ai/langchain

★ 143,6KPython

RAGFlow

infiniflow/ragflow

★ 86,7KGo