Skip to main content
With W&B Weave, the Trace view Call panel can render inputs and outputs as a Chat view (message bubbles) when the logged data matches a chat-completion style shape. Integrations that patch OpenAI or Anthropic clients usually produce this shape automatically. If you log calls yourself, for example through the Weave Service API, you need to format request and response payloads so Weave can recognize them.

​
When the Chat view appears

Weave looks for patterns similar to common LLM chat APIs:
  • Inputs include a messages array. Each item should include role (for example user, assistant, or system) and content. Including a model field alongside messages matches what many providers log for a single completion call.
  • Outputs include a choices array. Each item should include a message object with a content field for the assistant reply.
If the data does not match this shape, the Call panel still shows Input and Output, but not the Chat layout.

​
Shape inputs and outputs

The following example shows the minimal structure for one user turn and one assistant reply. Values are illustrative. 
{
  "inputs": {
    "model": "gpt-4o",
    "messages": [
      { "role": "user", "content": "Why is the sky blue?" }
    ]
  },
  "output": {
    "choices": [
      {
        "message": {
          "content": "It is largely due to Rayleigh scattering of sunlight in the atmosphere."
        }
      }
    ]
  }
}
For multi-turn conversations, add more objects to messages with the appropriate role and content values.

​
Log calls with the Service API

To start and end calls over HTTP, send payloads that put the fields above on the call record:
  • On call start, set inputs to include messages (and typically model).
  • On call end, set output to include choices with message.content for the completion text.
A full walkthrough with working Python examples is in Use the Service API to trace. See the simple trace section for a minimal end-to-end sample.

​
Log calls with the SDK or integrations

When you use W&B Weave integrations with provider SDKs, traced calls often inherit the correct structure without manual formatting. For background on how chat-style data appears in thread views, see Chat view behavior in the Threads documentation.

​
Customize or override display

If you need different rendering than the default Chat view, you can still control how data appears using weave.Markdown and postprocessors. See View and customize trace display.

​
OpenTelemetry traces

Traces ingested as OpenTelemetry may not show tool calls in the Chat view the same way as native Weave calls. For details, see Send OpenTelemetry traces to Weave.