Weave
The GenAI development toolkit for tracing, evaluating, and protecting LLM and generative AI applications.
Traces every call in your LLM app (prompts, responses, tokens, cost, latency) and quantifies quality with dataset-based evaluation (LLM-as-a-judge). In production, guardrails and online evaluations monitor safety and quality in real time.
What Is Weave
Generative AI apps produce non-deterministic output, making it hard to verify by eye whether they're working well. W&B Weave is a GenAI-specific development and observability tool that traces LLM application executions, quantifies quality through systematic evaluation, and protects production with guardrails. If Models is for training, Weave is for LLM apps.
Key Capabilities
Tracing
Traces prompts, responses, tool calls, tokens, cost, and latency as a tree
Evaluations
Quality evaluation with datasets plus LLM-as-a-judge and custom scorers
Playground
Experiment with multiple models and prompts side by side
Guardrails · Online Evals
PII and toxicity filters with real-time quality monitoring
Models vs Weave
| Aspect | W&B Models | W&B Weave |
|---|---|---|
| Target | ML training & models | LLM & generative AI apps |
| Core | Experiments · sweeps · registry | Tracing · evaluation · guardrails |
| Units | run · model · artifact | trace · evaluation · dataset |
What It's Used For
Debugging LLM apps
Trace failed calls to analyze root causes
Preventing quality regressions
Catch quality degradation early with pre-deployment evaluation
Production safety
Maintain safety and quality with guardrails and online evaluation