Capability · Framework — observability

W&B Weave

Weave is W&B's purpose-built product for generative-AI workflows, alongside their classic ML experiment-tracking tools. A single `@weave.op` decorator is enough to capture inputs, outputs, and latency for every call in your chain — whether it's an LLM, a tool, or a retrieval step. Evaluations and datasets layer on top, tightly integrated with W&B's experiment dashboard.

Framework facts

Category: observability
Language: Python / TypeScript
License: Apache-2.0 (SDK) / SaaS
Repository: https://github.com/wandb/weave

Install

pip install weave
wandb login

Quickstart

import weave
from openai import OpenAI

weave.init('my-llm-app')

@weave.op
def answer(q: str) -> str:
    resp = OpenAI().chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role':'user','content': q}]
    )
    return resp.choices[0].message.content

print(answer('hello'))
# see trace at https://wandb.ai/your-entity/my-llm-app/weave

Alternatives

Arize Phoenix — open-source equivalent
LangSmith — LangChain-native alternative
Langfuse — self-hostable open-source
Humanloop — prompt-IDE first

Frequently asked questions

Is Weave only for LLM apps?

It's optimised for LLMs and agents but tracks any Python function. Teams frequently use Weave to observe traditional ML scoring pipelines alongside generative calls.

Is the SDK open-source?

Yes — the SDK is Apache-2.0 on GitHub. The storage, dashboards, and collaboration features live in the hosted W&B platform (free and paid tiers available).

Sources

Weave — docs — accessed 2026-04-20
Weave GitHub — accessed 2026-04-20