# Weave — Trace and Debug LLM Apps

> Weave adds tracing to LLM apps with `@weave.op`. Install `weave`, call `weave.init()`, then track inputs/outputs across API calls and validation steps.

## Install

Copy the content below into your project:

# Weave — Trace and Debug LLM Apps

> Weave adds tracing to LLM apps with `@weave.op`. Install `weave`, call `weave.init()`, then track inputs/outputs across API calls and validation steps.

## Quick Use

1. Install:
   ```bash
   pip install weave
   ```
2. Run:
   ```bash
   python -c "import weave; weave.init('tokrepo-demo'); print('weave init ok')"
   ```
3. Verify:
   - Wrap a function with `@weave.op`, run it once, and confirm a trace is recorded


---

## Intro

Weave adds tracing to LLM apps with `@weave.op`. Install `weave`, call `weave.init()`, then track inputs/outputs across API calls and validation steps.

- **Best for:** teams debugging agent workflows who need end-to-end traces across tool calls, validation ops, and model calls
- **Works with:** Python, `weave.op` decorator, integration with LLM API calls and custom validation functions (repo examples)
- **Setup time:** 9 minutes


### Quantitative Notes

- Install: `pip install weave` (repo)
- Setup time ~9 minutes
- GitHub stars (verified): see Source & Thanks


---

## Practical Notes

Weave is most valuable when you treat observability as a product feature. Add traces around every boundary: user input → prompt assembly → tool calls → validation → final output. Then use the trace tree to answer: where did latency come from, which step failed, and what data caused the failure?

**Safety note:** Avoid logging secrets; sanitize prompts and tool args before tracing in production.

### FAQ

**Q: What should I trace first?**
A: Start with your top-level agent run function and the model-call wrapper; then add ops around tools and validators.

**Q: Does it only work with OpenAI?**
A: No. The README says you can trace any function, including calls to different providers and open-source models.

**Q: How do I keep traces useful?**
A: Log structured inputs, normalize outputs, and capture error cases; traces are only helpful when they show failures clearly.

---

## Source & Thanks

> GitHub: https://github.com/wandb/weave
> Owner avatar: https://avatars.githubusercontent.com/u/26401354?v=4
> License (SPDX): Apache-2.0
> GitHub stars (verified via `api.github.com/repos/wandb/weave`): 1,090


---

<!-- ZH -->

# Weave——为 LLM 应用做 Trace 与调试

> Weave 用 `@weave.op` 给 LLM 应用加 tracing。安装 `weave` 后调用 `weave.init()`，即可跟踪 API 调用与校验步骤的输入/输出，便于调试与回放。

## 快速使用

1. 安装：
   ```bash
   pip install weave
   ```
2. 运行：
   ```bash
   python -c "import weave; weave.init('tokrepo-demo'); print('weave init ok')"
   ```
3. 验证：
   - Wrap a function with `@weave.op`, run it once, and confirm a trace is recorded


---

## 简介

Weave 用 `@weave.op` 给 LLM 应用加 tracing。安装 `weave` 后调用 `weave.init()`，即可跟踪 API 调用与校验步骤的输入/输出，便于调试与回放。

- **适合谁（Best for）:** 在调试 agent 工作流的团队，需要跨工具调用、校验函数与模型调用的端到端 trace
- **兼容工具（Works with）:** Python、`weave.op` 装饰器、与 LLM API 调用/自定义校验函数结合（仓库示例）
- **安装时间（Setup time）:** 9 分钟


### 量化信息

- 安装：`pip install weave`（仓库）
- 装机约 9 分钟
- GitHub stars（已核验）：见「来源与感谢」


---

## 实战要点

把可观测性当成产品能力来做时，Weave 的价值最大。把 trace 放在每个边界上：用户输入 → prompt 组装 → 工具调用 → 校验 → 最终输出。用 trace 树回答三个问题：延迟来自哪里、哪一步失败、失败由什么数据触发。

**安全提示：** 生产环境避免记录敏感信息；在 trace 前先对 prompt 与工具参数做脱敏。

### FAQ

**Q: 优先 trace 哪些函数？**
A: 先包最外层的 agent run 和模型调用封装，再逐步把工具调用与校验函数也纳入 tracing。

**Q: 只能配 OpenAI 吗？**
A: 不是。README 表示可 trace 任意函数，包括不同 provider 的 API 调用或开源模型推理。

**Q: 怎样让 trace 真有用？**
A: 记录结构化输入、规范化输出，并重点捕获错误分支；只有能清楚呈现失败，trace 才有价值。

---

## 来源与感谢

> GitHub：https://github.com/wandb/weave
> Owner avatar：https://avatars.githubusercontent.com/u/26401354?v=4
> 许可证（SPDX）：Apache-2.0
> GitHub stars（已通过 `api.github.com/repos/wandb/weave` 核验）：1,090


---
Source: https://tokrepo.com/en/workflows/weave-trace-and-debug-llm-apps
Author: Agent Toolkit