weave-integration（编织整合）

Weave Integration

Add W&B Weave observability to any LLM project. Weave traces all LLM calls, captures tokens/latency/costs, and provides a UI for debugging and evaluation.

Quick Start

1. Install

# TypeScript/Node.js
npm install weave

# Python
pip install weave

2. Get API Key

Set WANDB_API_KEY environment variable. Get key from wandb.ai/settings.

export WANDB_API_KEY="your-key-here"

3. Initialize

// TypeScript
import * as weave from 'weave';
await weave.init('your-team/project-name');

# Python
import weave
weave.init('your-team/project-name')

4. Trace LLM Calls

Auto-patching (supported providers traced automatically):

// TypeScript - CommonJS: works out of the box
import OpenAI from 'openai';
import * as weave from 'weave';

await weave.init('my-project');
const client = new OpenAI();
// All calls now traced automatically

Manual wrapping (for custom functions or unsupported libs):

// TypeScript
const myFunction = weave.op(async (input: string) => {
  // your code here
  return result;
});

# Python
@weave.op()
def my_function(input: str):
    return result

TypeScript Setup Details

See references/typescript.md for:

• ESM configuration (--import=weave/instrument)
• Bundler compatibility (Next.js, Vite)
• Manual patching fallback

Supported Providers (Auto-traced)

OpenAI, Anthropic, Cohere, Mistral, Google, Groq, Together AI, LiteLLM, Azure, Bedrock, Cerebras, HuggingFace, OpenRouter, NVIDIA NIM, and more.

Full list: https://docs.wandb.ai/weave/guides/integrations

Integration Workflow

When adding Weave to a project:

1. Find LLM call sites — search for OpenAI/Anthropic client usage
2. Add weave.init() — early in app startup, before any LLM calls
3. Verify auto-patching — check traces appear in W&B UI
4. Wrap custom functions — use weave.op() for additional visibility
5. Add cost tracking — Weave tracks tokens automatically for supported providers

Viewing Traces

After running your app:

• Open wandb.ai → Your project → Weave tab
• See all traces with inputs, outputs, latency, token usage, costs
• Filter, search, and export call data

Environment Variables

Common Patterns

Wrap Existing Client

import { wrapOpenAI } from 'weave';
import OpenAI from 'openai';

const client = wrapOpenAI(new OpenAI());

Trace Class Methods

class MyAgent {
  @weave.op
  async predict(prompt: string) {
    return "response";
  }
}

Add Display Names

const myOp = weave.op(myFunction, {
  callDisplayName: (input) => `Custom Name: ${input}`
});

Clawdbot-Specific Integration

For Clawdbot/similar Node.js agents:

1. Locate the LLM client initialization (usually Anthropic/OpenAI SDK)
2. Add weave.init() in the main entry point
3. For ESM, add --import=weave/instrument to node invocation
4. All provider calls will be traced to W&B

Troubleshooting

• No traces appearing: Check WANDB_API_KEY is set
• ESM not patching: Use --import=weave/instrument flag
• Bundler issues: Mark LLM libs as external in config
• Manual fallback: Use wrapOpenAI() or explicit weave.op() wrappers