Send OpenTelemetry Traces to Weave

Ingest OpenTelemetry compatible trace data through a dedicated endpoint

Overview

Weave supports ingestion of OpenTelemetry compatible trace data through a dedicated endpoint. This endpoint allows you to send OTLP (OpenTelemetry Protocol) formatted trace data directly to your Weave project.

Endpoint details

Path: /otel/v1/traces Method: POST Content-Type: application/x-protobuf Base URL: The base URL for the OTel trace endpoint depends on your W&B deployment type:

• Multi-tenant Cloud:
  https://trace.wandb.ai/otel/v1/traces

• Dedicated Cloud and Self-Managed instances:
  https://<your-subdomain>.wandb.io/traces/otel/v1/traces

Replace <your-subdomain> with your organization’s unique W&B domain, e.g., acme.wandb.io.

Authentication and routing

Pass your W&B API key in the wandb-api-key header, then specify the following keys as OpenTelemetry Resource attributes in your TracerProvider class:

• wandb.entity: Your W&B team or user name.
• wandb.project: The project name to send traces to.

The following example shows how to configure authentication and project routing:

Python

import os
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.resources import Resource

WANDB_BASE_URL = "https://trace.wandb.ai"
ENTITY = "<your-team-name>"
PROJECT = "<your-project-name>"

OTEL_EXPORTER_OTLP_ENDPOINT = f"{WANDB_BASE_URL}/otel/v1/traces"

# Create an API key at https://wandb.ai/settings
WANDB_API_KEY = os.environ["WANDB_API_KEY"]

exporter = OTLPSpanExporter(
    endpoint=OTEL_EXPORTER_OTLP_ENDPOINT,
    headers={"wandb-api-key": WANDB_API_KEY},
)

tracer_provider = trace_sdk.TracerProvider(resource=Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
}))

import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { Resource } from "@opentelemetry/resources";

const WANDB_BASE_URL = "https://trace.wandb.ai";
const ENTITY = "<your-team-name>";
const PROJECT = "<your-project-name>";

const OTEL_EXPORTER_OTLP_ENDPOINT = `${WANDB_BASE_URL}/otel/v1/traces`;

// Create an API key at https://wandb.ai/settings
const WANDB_API_KEY = process.env.WANDB_API_KEY!;

const exporter = new OTLPTraceExporter({
  url: OTEL_EXPORTER_OTLP_ENDPOINT,
  headers: { "wandb-api-key": WANDB_API_KEY },
});

const provider = new NodeTracerProvider({
  resource: new Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
  }),
});

Examples

The following examples show how to send OpenTelemetry traces to Weave using Python and TypeScript.

Before running the code samples below, set the following fields:

1. WANDB_API_KEY: You can get this from User Settings.
2. Entity: You can only log traces to the project under an entity that you have access to. You can find your entity name by visiting your W&B dashboard at [https://wandb.ai/home], and checking the Teams field in the left sidebar.
3. Project Name: Choose a fun name!
4. OPENAI_API_KEY: You can obtain this from the OpenAI dashboard.

OpenInference Instrumentation

This example shows how to use the OpenAI instrumentation. There are many more available which you can find in the official repository.

First, install the required dependencies:

Python

pip install openai openinference-instrumentation-openai opentelemetry-exporter-otlp-proto-http

TypeScript

npm install openai @opentelemetry/sdk-trace-node @opentelemetry/sdk-trace-base @opentelemetry/resources @opentelemetry/exporter-trace-otlp-proto @arizeai/openinference-instrumentation-openai @opentelemetry/api

Performance Recommendation: Always use BatchSpanProcessor instead of SimpleSpanProcessor when sending traces to Weave. SimpleSpanProcessor exports spans synchronously, potentially impacting the performance of other workloads. These examples illustrate BatchSpanProcessor, which is recommended in production because it batches spans asynchronously and efficiently.

Python

Paste the following code into a Python file such as openinference_example.py:

import os
import openai
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, BatchSpanProcessor
from openinference.instrumentation.openai import OpenAIInstrumentor

OPENAI_API_KEY = "YOUR_OPENAI_API_KEY"
WANDB_BASE_URL = "https://trace.wandb.ai"
ENTITY = "<your-team-name>"
PROJECT = "<your-project-name>"

OTEL_EXPORTER_OTLP_ENDPOINT = f"{WANDB_BASE_URL}/otel/v1/traces"

# Create an API key at https://wandb.ai/settings
WANDB_API_KEY = os.environ["WANDB_API_KEY"]

exporter = OTLPSpanExporter(
    endpoint=OTEL_EXPORTER_OTLP_ENDPOINT,
    headers={"wandb-api-key": WANDB_API_KEY},
)

tracer_provider = trace_sdk.TracerProvider(resource=Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
}))
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))

# Optionally, print the spans to the console.
tracer_provider.add_span_processor(BatchSpanProcessor(ConsoleSpanExporter()))

OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)

def main():
    client = openai.OpenAI(api_key=OPENAI_API_KEY)
    response = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[{"role": "user", "content": "Describe OTel in a single sentence."}],
        max_tokens=20,
        stream=True,
        stream_options={"include_usage": True},
    )
    for chunk in response:
        if chunk.choices and (content := chunk.choices[0].delta.content):
            print(content, end="")

if __name__ == "__main__":
    main()

Run the code:

python openinference_example.py

TypeScript

The TypeScript implementation of this example contains the following key differences from the Python implementation:

• OpenAI must be imported before registering instrumentation (ESM modules require this).
• Uses @opentelemetry/exporter-trace-otlp-proto (protobuf format) instead of the HTTP exporter, since W&B’s endpoint only accepts protobuf.
• Requires explicit provider.shutdown() with a delay before shutdown to ensure spans are flushed, since BatchSpanProcessor flushes asynchronously.

Paste the following code into a TypeScript file such as openinference_example.ts:

// IMPORTANT: Import OpenAI FIRST so instrumentation can patch it
import OpenAI from "openai";
import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { BatchSpanProcessor } from "@opentelemetry/sdk-trace-base";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { resourceFromAttributes } from "@opentelemetry/resources";
import { OpenAIInstrumentation, isPatched } from "@arizeai/openinference-instrumentation-openai";

const OPENAI_API_KEY = process.env.OPENAI_API_KEY;
const WANDB_BASE_URL = "https://trace.wandb.ai";
const ENTITY = "<your-team-name>";
const PROJECT = "<your-project-name>";

const OTEL_EXPORTER_OTLP_ENDPOINT = `${WANDB_BASE_URL}/otel/v1/traces`;

// Create an API key at https://wandb.ai/settings
const WANDB_API_KEY = process.env.WANDB_API_KEY!;

const exporter = new OTLPTraceExporter({
  url: OTEL_EXPORTER_OTLP_ENDPOINT,
  headers: { "wandb-api-key": WANDB_API_KEY },
});

const provider = new NodeTracerProvider({
  resource: resourceFromAttributes({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
  }),
  spanProcessors: [
    new BatchSpanProcessor(exporter)
  ],
});

provider.register();

// Register the OpenAI instrumentation with the tracer provider
const openAIInstrumentation = new OpenAIInstrumentation();
openAIInstrumentation.setTracerProvider(provider);

// Manually instrument OpenAI since we're using ESM
openAIInstrumentation.manuallyInstrument(OpenAI);

async function main() {
  console.log("OpenAI is patched?", isPatched());

  const client = new OpenAI({ apiKey: OPENAI_API_KEY });

  console.log("Making OpenAI API call...");
  const response = await client.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: [{ role: "user", content: "Describe OTel in a single sentence." }],
    max_tokens: 50,
  });

  console.log("Response:", response.choices[0]?.message?.content);
  console.log("Waiting for spans to flush...");
}

(async () => {
  await main();

  // Give spans time to flush
  console.log("Waiting 2 seconds for spans to flush...");
  await new Promise(resolve => setTimeout(resolve, 2000));

  await provider.shutdown(); // flush all pending spans before exit
  console.log("Shutdown complete");
})();

Run the code:

npx ts-node openinference_example.ts

OpenLLMetry Instrumentation

The following example shows how to use the OpenAI instrumentation. Additional examples are available in the OpenLLMetry repository.

First, install the required dependencies:

Python

pip install openai opentelemetry-instrumentation-openai opentelemetry-exporter-otlp-proto-http

TypeScript

npm install openai @traceloop/instrumentation-openai @opentelemetry/sdk-trace-node @opentelemetry/resources @opentelemetry/exporter-trace-otlp-http

Python

Paste the following code into a Python file such as openllmetry_example.py. Note that this is the same code as above, except the OpenAIInstrumentor is imported from opentelemetry.instrumentation.openai instead of openinference.instrumentation.openai:

import os
import openai
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, BatchSpanProcessor
from opentelemetry.instrumentation.openai import OpenAIInstrumentor

OPENAI_API_KEY = "YOUR_OPENAI_API_KEY"
WANDB_BASE_URL = "https://trace.wandb.ai"
ENTITY = "<your-team-name>"
PROJECT = "<your-project-name>"

OTEL_EXPORTER_OTLP_ENDPOINT = f"{WANDB_BASE_URL}/otel/v1/traces"

# Create an API key at https://wandb.ai/settings
WANDB_API_KEY = os.environ["WANDB_API_KEY"]

exporter = OTLPSpanExporter(
    endpoint=OTEL_EXPORTER_OTLP_ENDPOINT,
    headers={"wandb-api-key": WANDB_API_KEY},
)

tracer_provider = trace_sdk.TracerProvider(resource=Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
}))
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))

# Optionally, print the spans to the console.
tracer_provider.add_span_processor(BatchSpanProcessor(ConsoleSpanExporter()))

OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)

def main():
    client = openai.OpenAI(api_key=OPENAI_API_KEY)
    response = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[{"role": "user", "content": "Describe OTel in a single sentence."}],
        max_tokens=20,
        stream=True,
        stream_options={"include_usage": True},
    )
    for chunk in response:
        if chunk.choices and (content := chunk.choices[0].delta.content):
            print(content, end="")

if __name__ == "__main__":
    main()

Run the code:

python openllmetry_example.py

TypeScript

Paste the following code into a TypeScript file such as openllmetry_example.ts. Note that this uses the Traceloop OpenAI instrumentation package:

import OpenAI from "openai";
import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { BatchSpanProcessor, ConsoleSpanExporter } from "@opentelemetry/sdk-trace-base";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { Resource } from "@opentelemetry/resources";
import { OpenAIInstrumentation } from "@traceloop/instrumentation-openai";
import { registerInstrumentations } from "@opentelemetry/instrumentation";

const OPENAI_API_KEY = process.env.OPENAI_API_KEY;
const WANDB_BASE_URL = "https://trace.wandb.ai";
const ENTITY = "<your-team-name>";
const PROJECT = "<your-project-name>";

const OTEL_EXPORTER_OTLP_ENDPOINT = `${WANDB_BASE_URL}/otel/v1/traces`;

// Create an API key at https://wandb.ai/settings
const WANDB_API_KEY = process.env.WANDB_API_KEY!;

const exporter = new OTLPTraceExporter({
  url: OTEL_EXPORTER_OTLP_ENDPOINT,
  headers: { "wandb-api-key": WANDB_API_KEY },
});

const provider = new NodeTracerProvider({
  resource: new Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
  }),
  spanProcessors: [
    new BatchSpanProcessor(exporter),
    // Optionally, print the spans to the console.
    new BatchSpanProcessor(new ConsoleSpanExporter()),
  ],
});

provider.register();

// Register the OpenAI instrumentation with the tracer provider
const openAIInstrumentation = new OpenAIInstrumentation();
registerInstrumentations({
  tracerProvider: provider,
  instrumentations: [openAIInstrumentation],
});

// Manually instrument OpenAI since we're using ESM
openAIInstrumentation.manuallyInstrument(OpenAI);

async function main() {
  const client = new OpenAI({ apiKey: OPENAI_API_KEY });
  const stream = await client.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: [{ role: "user", content: "Describe OTel in a single sentence." }],
    max_tokens: 20,
    stream: true,
  });

  for await (const chunk of stream) {
    const content = chunk.choices[0]?.delta?.content;
    if (content) {
      process.stdout.write(content);
    }
  }
  console.log(); // newline after streaming
}

(async () => {
  await main();

  // Give spans time to flush
  await new Promise(resolve => setTimeout(resolve, 2000));

  await provider.shutdown(); // flush all pending spans before exit
})();

Run the code:

npx ts-node openllmetry_example.ts

Without instrumentation

If you would prefer to use OTel directly instead of an instrumentation package, you may do so. Span attributes will be parsed according to the OpenTelemetry semantic conventions described at https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/.

First, install the required dependencies:

Python

pip install openai opentelemetry-sdk opentelemetry-api opentelemetry-exporter-otlp-proto-http

TypeScript

npm install openai @opentelemetry/api @opentelemetry/sdk-trace-node @opentelemetry/resources @opentelemetry/exporter-trace-otlp-http

Python

Paste the following code into a Python file such as opentelemetry_example.py:

import json
import os
import openai
from opentelemetry import trace
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.resources import Resource
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, BatchSpanProcessor

OPENAI_API_KEY = "YOUR_OPENAI_API_KEY"
WANDB_BASE_URL = "https://trace.wandb.ai"
ENTITY = "<your-team-name>"
PROJECT = "<your-project-name>"

OTEL_EXPORTER_OTLP_ENDPOINT = f"{WANDB_BASE_URL}/otel/v1/traces"

# Create an API key at https://wandb.ai/settings
WANDB_API_KEY = os.environ["WANDB_API_KEY"]

# Configure the OTLP exporter
exporter = OTLPSpanExporter(
    endpoint=OTEL_EXPORTER_OTLP_ENDPOINT,
    headers={"wandb-api-key": WANDB_API_KEY},
)

tracer_provider = trace_sdk.TracerProvider(resource=Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
}))
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))

# Optionally, print the spans to the console.
tracer_provider.add_span_processor(BatchSpanProcessor(ConsoleSpanExporter()))

# Set the tracer provider
trace.set_tracer_provider(tracer_provider)

# Create a tracer from the global tracer provider
tracer = trace.get_tracer(__name__)

def my_function():
    with tracer.start_as_current_span("outer_span") as outer_span:
        client = openai.OpenAI()
        input_messages = [{"role": "user", "content": "Describe OTel in a single sentence."}]
        outer_span.set_attribute("input.value", json.dumps(input_messages))
        outer_span.set_attribute("gen_ai.system", "openai")
        response = client.chat.completions.create(
            model="gpt-3.5-turbo",
            messages=input_messages,
            max_tokens=20,
            stream=True,
            stream_options={"include_usage": True},
        )
        out = ""
        for chunk in response:
            if chunk.choices and (content := chunk.choices[0].delta.content):
                out += content
        outer_span.set_attribute("output.value", json.dumps({"content": out}))

if __name__ == "__main__":
    my_function()

Run the code:

python opentelemetry_example.py

TypeScript

Paste the following code into a TypeScript file such as opentelemetry_example.ts:

import OpenAI from "openai";
import { trace } from "@opentelemetry/api";
import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { BatchSpanProcessor, ConsoleSpanExporter } from "@opentelemetry/sdk-trace-base";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-http";
import { Resource } from "@opentelemetry/resources";

const OPENAI_API_KEY = "YOUR_OPENAI_API_KEY";
const WANDB_BASE_URL = "https://trace.wandb.ai";
const ENTITY = "<your-team-name>";
const PROJECT = "<your-project-name>";

const OTEL_EXPORTER_OTLP_ENDPOINT = `${WANDB_BASE_URL}/otel/v1/traces`;

// Create an API key at https://wandb.ai/settings
const WANDB_API_KEY = process.env.WANDB_API_KEY!;

const exporter = new OTLPTraceExporter({
  url: OTEL_EXPORTER_OTLP_ENDPOINT,
  headers: { "wandb-api-key": WANDB_API_KEY },
});

const provider = new NodeTracerProvider({
  resource: new Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
  }),
  spanProcessors: [
    new BatchSpanProcessor(exporter),
    // Optionally, print the spans to the console.
    new BatchSpanProcessor(new ConsoleSpanExporter()),
  ],
});

provider.register();

// Creates a tracer from the global tracer provider
const tracer = trace.getTracer("my-app");

async function myFunction() {
  const span = tracer.startSpan("outer_span");

  try {
    const client = new OpenAI({ apiKey: OPENAI_API_KEY });
    const inputMessages = [
      { role: "user" as const, content: "Describe OTel in a single sentence." },
    ];

    // This will only appear in the side panel
    span.setAttribute("input.value", JSON.stringify(inputMessages));
    
    // This follows conventions and will appear in the dashboard
    span.setAttribute("gen_ai.system", "openai");

    const stream = await client.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: inputMessages,
      max_tokens: 20,
      stream: true,
    });

    let output = "";
    for await (const chunk of stream) {
      const content = chunk.choices[0]?.delta?.content;
      if (content) {
        output += content;
      }
      }

    // This will only appear in the side panel
    span.setAttribute("output.value", JSON.stringify({ content: output }));
  } finally {
    span.end();
  }
}

myFunction();

Run the code:

npx ts-node opentelemetry_example.ts

The span attribute prefixes gen_ai and openinference are used to determine which convention to use, if any, when interpreting the trace. If neither key is detected, then all span attributes are visible in the trace view. The full span is available in the side panel when you select a trace.

Use an OpenTelemetry Collector

The examples above export traces directly from your application to Weave. In production, you can use an OpenTelemetry Collector as an intermediary between your application and Weave. The collector receives traces from your app, then forwards them to one or more backends.

Set up a collector

The following example shows how to:

• Set up a Docker configuration file that deploys a local server (collector) that listens for OTLP traces, batches them, and forwards them to Weave.
• Locally run the collector using Docker.
• Send a basic call to OpenAI that forwards traces to the collector running in the Docker container.

To use a collector, first create a collector-config.yaml file that configures the collector to receive OTLP traces and export them to Weave:

collector-config.yaml

receivers:
  otlp:
    protocols:
      http:
        endpoint: 0.0.0.0:4318

exporters:
  otlphttp/weave:
    endpoint: ${env:WANDB_OTLP_ENDPOINT}
    headers:
      wandb-api-key: ${env:WANDB_API_KEY}
    sending_queue:
      batch:

processors:
  resource:
    attributes:
      - key: wandb.entity # Resource attributes field
        value: ${env:DEFAULT_WANDB_ENTITY}  # Value to inject
        action: insert # Inject only if not already present
      - key: wandb.project
        value: ${env:DEFAULT_WANDB_PROJECT}
        action: insert 

service:
  pipelines:
    traces:
      receivers: [otlp]
      processors: [resource]
      exporters: [otlphttp/weave]

This configuration file:

• Listens for OTLP traces on port 4318 (HTTP).
• Exports traces to Weave’s OTLP endpoint using the wandb-api-key header, reading the endpoint URL from WANDB_OTLP_ENDPOINT and the API key from WANDB_API_KEY.
• Sets wandb.entity and wandb.project as resource attributes using the resource processor, reading values from DEFAULT_WANDB_ENTITY and DEFAULT_WANDB_PROJECT. The insert action injects these attributes only if your application code does not already set them.
• Enables the exporter’s built-in sending_queue with batching to reduce network overhead.

After configuring the collector’s settings, update the API and entity values in the following Docker command and run it:

docker run \
  -v ./config.yaml:/etc/otelcol-contrib/config.yaml \
  -e WANDB_API_KEY="<your-wandb-api-key>" \
  -e WANDB_OTLP_ENDPOINT="https://trace.wandb.ai/otel" \
  -e DEFAULT_WANDB_ENTITY="<your-team-name>" \
  -e DEFAULT_WANDB_PROJECT="YOUR_PROJECT" \
  -p 4318:4318 \
  otel/opentelemetry-collector-contrib:latest

Once the collector is running, configure your application to export traces to it by setting the OTEL_EXPORTER_OTLP_ENDPOINT environment variable. The OTel SDK reads this variable automatically, so you do not need to pass the endpoint to the exporter.

If you set wandb.entity or wandb.project as resource attributes in your application’s TracerProvider, they take precedence over the defaults defined in the collector config.

Python

import os
import openai
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from openinference.instrumentation.openai import OpenAIInstrumentor

os.environ["OTEL_EXPORTER_OTLP_ENDPOINT"] = "http://localhost:4318"

OPENAI_API_KEY = "YOUR_OPENAI_API_KEY"

tracer_provider = trace_sdk.TracerProvider()
tracer_provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))

OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)

def main():
    client = openai.OpenAI(api_key=OPENAI_API_KEY)
    response = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[{"role": "user", "content": "Describe OTel in a single sentence."}],
        max_tokens=20,
    )
    print(response.choices[0].message.content)

if __name__ == "__main__":
    main()

import OpenAI from "openai";
import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { BatchSpanProcessor } from "@opentelemetry/sdk-trace-base";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { OpenAIInstrumentation, isPatched } from "@arizeai/openinference-instrumentation-openai";

process.env.OTEL_EXPORTER_OTLP_ENDPOINT = "http://localhost:4318";

const OPENAI_API_KEY = process.env.OPENAI_API_KEY;

const provider = new NodeTracerProvider({
  spanProcessors: [new BatchSpanProcessor(new OTLPTraceExporter())],
});

provider.register();

const openAIInstrumentation = new OpenAIInstrumentation();
openAIInstrumentation.setTracerProvider(provider);
openAIInstrumentation.manuallyInstrument(OpenAI);

async function main() {
  console.log("OpenAI is patched?", isPatched());

  const client = new OpenAI({ apiKey: OPENAI_API_KEY });
  const response = await client.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: [{ role: "user", content: "Describe OTel in a single sentence." }],
    max_tokens: 20,
  });

  console.log("Response:", response.choices[0]?.message?.content);
}

(async () => {
  await main();
  await new Promise(resolve => setTimeout(resolve, 2000));
  await provider.shutdown();
})();

The OpenAIInstrumentor automatically wraps OpenAI calls, creates traces, and exports them to the collector. The collector handles authentication and routing to Weave.

After running the script, you can view the traces in the Weave UI.

To send traces to additional backends, add more exporters and include them in the service.pipelines.traces.exporters list. For example, you can export to both Weave and Jaeger from the same Collector instance.

Organize OTel traces into threads

Add specific span attributes to organize your OpenTelemetry traces into Weave threads, then use Weave’s Thread UI to analyze related operations like multi-turn conversations or user sessions in Weave’s thread UI.

Add the following attributes to your OTel spans to enable thread grouping:

• wandb.thread_id: Groups spans into a specific thread
• wandb.is_turn: Marks a span as a conversation turn (appears as a row in the thread view)

The following examples show how to organize OTel traces into Weave threads. They use wandb.thread_id to group related operations and wandb.is_turn to mark high-level operations that appear as rows in the thread view.

Initial set up

Use this configuration to run these examples:

Python

import json
import os
from opentelemetry import trace
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.resources import Resource
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, BatchSpanProcessor

# Configuration
ENTITY = "<your-team-name>"
PROJECT = "<your-project-name>"
WANDB_API_KEY = os.environ["WANDB_API_KEY"]

OTEL_EXPORTER_OTLP_ENDPOINT = "https://trace.wandb.ai/otel/v1/traces"

exporter = OTLPSpanExporter(
    endpoint=OTEL_EXPORTER_OTLP_ENDPOINT,
    headers={"wandb-api-key": WANDB_API_KEY},
)

tracer_provider = trace_sdk.TracerProvider(resource=Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
}))
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))

# Optionally, print the spans to the console
tracer_provider.add_span_processor(BatchSpanProcessor(ConsoleSpanExporter()))

trace.set_tracer_provider(tracer_provider)

# Creates a tracer from the global tracer provider
tracer = trace.get_tracer(__name__)

TypeScript

import { trace, context } from "@opentelemetry/api";
import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import {
  BatchSpanProcessor,
  ConsoleSpanExporter,
} from "@opentelemetry/sdk-trace-base";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { Resource } from "@opentelemetry/resources";

// Configuration
const ENTITY = "<your-team-name>";
const PROJECT = "<your-project-name>";
const WANDB_API_KEY = process.env.WANDB_API_KEY;

if (!WANDB_API_KEY) {
  console.error("Error: WANDB_API_KEY environment variable is not set");
  console.error("Run: export WANDB_API_KEY=your_api_key_here");
  process.exit(1);
}

// OTel Setup
const OTEL_EXPORTER_OTLP_ENDPOINT = "https://trace.wandb.ai/otel/v1/traces";

const exporter = new OTLPTraceExporter({
  url: OTEL_EXPORTER_OTLP_ENDPOINT,
  headers: { "wandb-api-key": WANDB_API_KEY },
});

// Initialize tracer provider with span processors
const provider = new NodeTracerProvider({
  resource: new Resource({
    "wandb.entity": ENTITY,
    "wandb.project": PROJECT,
  }),
  spanProcessors: [
    new BatchSpanProcessor(exporter),
    new BatchSpanProcessor(new ConsoleSpanExporter()),
  ],
});

// Register the tracer provider
provider.register();

// Create a tracer from the global tracer provider
const tracer = trace.getTracer("threads-examples");

Trace a basic single-turn thread

Python

def example_1_basic_thread_and_turn():
    """Example 1: Basic thread with a single turn"""
    print("\n=== Example 1: Basic Thread and Turn ===")

    # Create a thread context
    thread_id = "thread_example_1"

    # This span represents a turn (direct child of thread)
    with tracer.start_as_current_span("process_user_message") as turn_span:
        # Set thread attributes
        turn_span.set_attribute("wandb.thread_id", thread_id)
        turn_span.set_attribute("wandb.is_turn", True)

        # Add some example attributes
        turn_span.set_attribute("input.value", "Hello, help me with setup")

        # Simulate some work with nested spans
        with tracer.start_as_current_span("generate_response") as nested_span:
            # This is a nested call within the turn, so is_turn should be false or unset
            nested_span.set_attribute("wandb.thread_id", thread_id)
            # wandb.is_turn is not set or set to False for nested calls

            response = "I'll help you get started with the setup process."
            nested_span.set_attribute("output.value", response)

        turn_span.set_attribute("output.value", response)
        print(f"Turn completed in thread: {thread_id}")

def main():
    example_1_basic_thread_and_turn()

if __name__ == "__main__":
    main()

TypeScript

function example_1_basic_thread_and_turn() {
  console.log("\n=== Example 1: Basic Thread and Turn ===");

  // Create a thread context
  const threadId = "thread_example_1";

  // This span represents a turn (direct child of thread)
  tracer.startActiveSpan("process_user_message", (turnSpan) => {
    // Set thread attributes
    turnSpan.setAttribute("wandb.thread_id", threadId);
    turnSpan.setAttribute("wandb.is_turn", true);

    // Add some example attributes
    turnSpan.setAttribute("input.value", "Hello, help me with setup");

    let response: string;
    
    // Simulate some work with nested spans
    tracer.startActiveSpan("generate_response", (nestedSpan) => {
      // This is a nested call within the turn, so is_turn should be false or unset
      nestedSpan.setAttribute("wandb.thread_id", threadId);
      // wandb.is_turn is not set or set to false for nested calls

      response = "I'll help you get started with the setup process.";
      nestedSpan.setAttribute("output.value", response);
      nestedSpan.end();
    });
    
    turnSpan.setAttribute("output.value", response!);
    console.log(`Turn completed in thread: ${threadId}`);
    turnSpan.end();
  });
}

function main() {
  example_1_basic_thread_and_turn();
}

main();

Trace a multi-turn conversation sharing one thread ID

Python

def example_2_multiple_turns():
    """Example 2: Multiple turns in a single thread"""
    print("\n=== Example 2: Multiple Turns in Thread ===")

    thread_id = "thread_conversation_123"

    # Turn 1
    with tracer.start_as_current_span("process_message_turn1") as turn1_span:
        turn1_span.set_attribute("wandb.thread_id", thread_id)
        turn1_span.set_attribute("wandb.is_turn", True)
        turn1_span.set_attribute("input.value", "What programming languages do you recommend?")

        # Nested operations
        with tracer.start_as_current_span("analyze_query") as analyze_span:
            analyze_span.set_attribute("wandb.thread_id", thread_id)
            # No is_turn attribute or set to False for nested spans

        response1 = "I recommend Python for beginners and JavaScript for web development."
        turn1_span.set_attribute("output.value", response1)
        print(f"Turn 1 completed in thread: {thread_id}")

    # Turn 2
    with tracer.start_as_current_span("process_message_turn2") as turn2_span:
        turn2_span.set_attribute("wandb.thread_id", thread_id)
        turn2_span.set_attribute("wandb.is_turn", True)
        turn2_span.set_attribute("input.value", "Can you explain Python vs JavaScript?")

        # Nested operations
        with tracer.start_as_current_span("comparison_analysis") as compare_span:
            compare_span.set_attribute("wandb.thread_id", thread_id)
            compare_span.set_attribute("wandb.is_turn", False)  # Explicitly false for nested

        response2 = "Python excels at data science while JavaScript dominates web development."
        turn2_span.set_attribute("output.value", response2)
        print(f"Turn 2 completed in thread: {thread_id}")

def main():
    example_2_multiple_turns()

if __name__ == "__main__":
    main()

TypeScript

function example_2_multiple_turns() {
  console.log("\n=== Example 2: Multiple Turns in Thread ===");

  const threadId = "thread_conversation_123";

  // Turn 1
  tracer.startActiveSpan("process_message_turn1", (turn1Span) => {
    turn1Span.setAttribute("wandb.thread_id", threadId);
    turn1Span.setAttribute("wandb.is_turn", true);
    turn1Span.setAttribute(
      "input.value",
      "What programming languages do you recommend?"
    );

    // Nested operations
    tracer.startActiveSpan("analyze_query", (analyzeSpan) => {
      analyzeSpan.setAttribute("wandb.thread_id", threadId);
      // No is_turn attribute or set to false for nested spans
      analyzeSpan.end();
    });

    const response1 =
      "I recommend Python for beginners and JavaScript for web development.";
    turn1Span.setAttribute("output.value", response1);
    console.log(`Turn 1 completed in thread: ${threadId}`);
    turn1Span.end();
  });

  // Turn 2
  tracer.startActiveSpan("process_message_turn2", (turn2Span) => {
    turn2Span.setAttribute("wandb.thread_id", threadId);
    turn2Span.setAttribute("wandb.is_turn", true);
    turn2Span.setAttribute("input.value", "Can you explain Python vs JavaScript?");

    // Nested operations
    tracer.startActiveSpan("comparison_analysis", (compareSpan) => {
      compareSpan.setAttribute("wandb.thread_id", threadId);
      compareSpan.setAttribute("wandb.is_turn", false); // Explicitly false for nested
      compareSpan.end();
    });

    const response2 =
      "Python excels at data science while JavaScript dominates web development.";
    turn2Span.setAttribute("output.value", response2);
    console.log(`Turn 2 completed in thread: ${threadId}`);
    turn2Span.end();
  });
}

function main() {
  example_2_multiple_turns();
}

main();

Trace deeply nested operations and mark only the outermost span as a turn

Python

def example_3_complex_nested_structure():
    """Example 3: Complex nested structure with multiple levels"""
    print("\n=== Example 3: Complex Nested Structure ===")

    thread_id = "thread_complex_456"

    # Turn with multiple levels of nesting
    with tracer.start_as_current_span("handle_complex_request") as turn_span:
        turn_span.set_attribute("wandb.thread_id", thread_id)
        turn_span.set_attribute("wandb.is_turn", True)
        turn_span.set_attribute("input.value", "Analyze this code and suggest improvements")

        # Level 1 nested operation
        with tracer.start_as_current_span("code_analysis") as analysis_span:
            analysis_span.set_attribute("wandb.thread_id", thread_id)
            # No is_turn for nested operations

            # Level 2 nested operation
            with tracer.start_as_current_span("syntax_check") as syntax_span:
                syntax_span.set_attribute("wandb.thread_id", thread_id)
                syntax_span.set_attribute("result", "No syntax errors found")

            # Another Level 2 nested operation
            with tracer.start_as_current_span("performance_check") as perf_span:
                perf_span.set_attribute("wandb.thread_id", thread_id)
                perf_span.set_attribute("result", "Found 2 optimization opportunities")

        # Another Level 1 nested operation
        with tracer.start_as_current_span("generate_suggestions") as suggest_span:
            suggest_span.set_attribute("wandb.thread_id", thread_id)
            suggestions = ["Use list comprehension", "Consider caching results"]
            suggest_span.set_attribute("suggestions", json.dumps(suggestions))

        turn_span.set_attribute("output.value", "Analysis complete with 2 improvement suggestions")
        print(f"Complex turn completed in thread: {thread_id}")

def main():
    example_3_complex_nested_structure()

if __name__ == "__main__":
    main()

TypeScript

function example_3_complex_nested_structure() {
  console.log("\n=== Example 3: Complex Nested Structure ===");

  const threadId = "thread_complex_456";

  // Turn with multiple levels of nesting
  tracer.startActiveSpan("handle_complex_request", (turnSpan) => {
    turnSpan.setAttribute("wandb.thread_id", threadId);
    turnSpan.setAttribute("wandb.is_turn", true);
    turnSpan.setAttribute(
      "input.value",
      "Analyze this code and suggest improvements"
    );

    // Level 1 nested operation
    tracer.startActiveSpan("code_analysis", (analysisSpan) => {
      analysisSpan.setAttribute("wandb.thread_id", threadId);
      // No is_turn for nested operations

      // Level 2 nested operation
      tracer.startActiveSpan("syntax_check", (syntaxSpan) => {
        syntaxSpan.setAttribute("wandb.thread_id", threadId);
        syntaxSpan.setAttribute("result", "No syntax errors found");
        syntaxSpan.end();
      });

      // Another Level 2 nested operation
      tracer.startActiveSpan("performance_check", (perfSpan) => {
        perfSpan.setAttribute("wandb.thread_id", threadId);
        perfSpan.setAttribute("result", "Found 2 optimization opportunities");
        perfSpan.end();
      });

      analysisSpan.end();
    });

    // Another Level 1 nested operation
    tracer.startActiveSpan("generate_suggestions", (suggestSpan) => {
      suggestSpan.setAttribute("wandb.thread_id", threadId);
      const suggestions = ["Use list comprehension", "Consider caching results"];
      suggestSpan.setAttribute("suggestions", JSON.stringify(suggestions));
      suggestSpan.end();
    });

    turnSpan.setAttribute(
      "output.value",
      "Analysis complete with 2 improvement suggestions"
    );
    console.log(`Complex turn completed in thread: ${threadId}`);
    turnSpan.end();
  });
}

function main() {
  example_3_complex_nested_structure();
}

main();

Trace background operations that belong to a thread but aren't turns

Python

def example_4_non_turn_operations():
    """Example 4: Operations that are part of a thread but not turns"""
    print("\n=== Example 4: Non-Turn Thread Operations ===")

    thread_id = "thread_background_789"

    # Background operation that's part of thread but not a turn
    with tracer.start_as_current_span("background_indexing") as bg_span:
        bg_span.set_attribute("wandb.thread_id", thread_id)
        # wandb.is_turn is unset or false - this is not a turn
        bg_span.set_attribute("wandb.is_turn", False)
        bg_span.set_attribute("operation", "Indexing conversation history")
        print(f"Background operation in thread: {thread_id}")

    # Actual turn in the same thread
    with tracer.start_as_current_span("user_query") as turn_span:
        turn_span.set_attribute("wandb.thread_id", thread_id)
        turn_span.set_attribute("wandb.is_turn", True)
        turn_span.set_attribute("input.value", "Search my previous conversations")
        turn_span.set_attribute("output.value", "Found 5 relevant conversations")
        print(f"Turn completed in thread: {thread_id}")

def main():
    example_4_non_turn_operations()

if __name__ == "__main__":
    main()

TypeScript

function example_4_non_turn_operations() {
  console.log("\n=== Example 4: Non-Turn Thread Operations ===");

  const threadId = "thread_background_789";

  // Background operation that's part of thread but not a turn
  tracer.startActiveSpan("background_indexing", (bgSpan) => {
    bgSpan.setAttribute("wandb.thread_id", threadId);
    // wandb.is_turn is unset or false - this is not a turn
    bgSpan.setAttribute("wandb.is_turn", false);
    bgSpan.setAttribute("operation", "Indexing conversation history");
    console.log(`Background operation in thread: ${threadId}`);
    bgSpan.end();
  });

  // Actual turn in the same thread
  tracer.startActiveSpan("user_query", (turnSpan) => {
    turnSpan.setAttribute("wandb.thread_id", threadId);
    turnSpan.setAttribute("wandb.is_turn", true);
    turnSpan.setAttribute("input.value", "Search my previous conversations");
    turnSpan.setAttribute("output.value", "Found 5 relevant conversations");
    console.log(`Turn completed in thread: ${threadId}`);
    turnSpan.end();
  });
}

function main() {
  example_4_non_turn_operations();
}

main();

After sending these traces, you can view them in the Weave UI under the Threads tab, where they’ll be grouped by thread_id and each turn will appear as a separate row.

Attribute mappings

Weave automatically maps OpenTelemetry span attributes from various instrumentation frameworks to its internal data model. When multiple attribute names map to the same field, Weave applies them in priority order, allowing frameworks to coexist in the same traces.

Supported frameworks

Weave supports attribute conventions from the following observability frameworks and SDKs:

• OpenTelemetry GenAI: Standard semantic conventions for generative AI (gen_ai.*)
• OpenInference: Arize AI’s instrumentation library (input.value, output.value, llm.*, openinference.*)
• Vercel AI SDK: Vercel’s AI SDK attributes (ai.prompt, ai.response, ai.model.*, ai.usage.*)
• MLflow: MLflow tracking attributes (mlflow.spanInputs, mlflow.spanOutputs)
• Traceloop: OpenLLMetry instrumentation (traceloop.entity.*, traceloop.span.kind)
• Google Vertex AI: Vertex AI agent attributes (gcp.vertex.agent.*)
• OpenLit: OpenLit observability attributes (gen_ai.content.completion)
• Langfuse: Langfuse tracing attributes (langfuse.startTime, langfuse.endTime)

Attribute reference

Attribute Field Name	W&B Mapping	Description	Type	Example
ai.prompt	inputs	User prompt text or messages.	String, list, dict	"Write a short haiku about summer."
gen_ai.prompt	inputs	AI model prompt or message array.	List, dict, string	[{"role":"user","content":"abc"}]
input.value	inputs	Input value for model invocation.	String, list, dict	{"text":"Tell a joke"}
mlflow.spanInputs	inputs	Span input data.	String, list, dict	["prompt text"]
traceloop.entity.input	inputs	Entity input data.	String, list, dict	"Translate this to French"
gcp.vertex.agent.tool_call_args	inputs	Tool call arguments.	Dict	{"args":{"query":"weather in SF"}}
gcp.vertex.agent.llm_request	inputs	LLM request payload.	Dict	{"contents":[{"role":"user","parts":[...]}]}
input	inputs	Generic input value.	String, list, dict	"Summarize this text"
inputs	inputs	Generic input array.	List, dict, string	["Summarize this text"]
ai.response	outputs	Model response text or data.	String, list, dict	"Here is a haiku..."
gen_ai.completion	outputs	AI completion result.	String, list, dict	"Completion text"
output.value	outputs	Output value from model.	String, list, dict	{"text":"Answer text"}
mlflow.spanOutputs	outputs	Span output data.	String, list, dict	["answer"]
gen_ai.content.completion	outputs	Content completion result.	String	"Answer text"
traceloop.entity.output	outputs	Entity output data.	String, list, dict	"Answer text"
gcp.vertex.agent.tool_response	outputs	Tool execution response.	Dict, string	{"toolResponse":"ok"}
gcp.vertex.agent.llm_response	outputs	LLM response payload.	Dict, string	{"candidates":[...]}
output	outputs	Generic output value.	String, list, dict	"Answer text"
outputs	outputs	Generic output array.	List, dict, string	["Answer text"]
gen_ai.usage.input_tokens	usage.input_tokens	Number of input tokens consumed.	Int	42
gen_ai.usage.prompt_tokens	usage.prompt_tokens	Number of prompt tokens consumed.	Int	30
llm.token_count.prompt	usage.prompt_tokens	Prompt token count.	Int	30
ai.usage.promptTokens	usage.prompt_tokens	Prompt tokens consumed.	Int	30
gen_ai.usage.completion_tokens	usage.completion_tokens	Number of completion tokens generated.	Int	40
llm.token_count.completion	usage.completion_tokens	Completion token count.	Int	40
ai.usage.completionTokens	usage.completion_tokens	Completion tokens generated.	Int	40
llm.usage.total_tokens	usage.total_tokens	Total tokens used in request.	Int	70
llm.token_count.total	usage.total_tokens	Total token count.	Int	70
gen_ai.system	attributes.system	System prompt or instructions.	String	"You are a helpful assistant."
llm.system	attributes.system	System prompt or instructions.	String	"You are a helpful assistant."
weave.span.kind	attributes.kind	Span type or category.	String	"llm"
traceloop.span.kind	attributes.kind	Span type or category.	String	"llm"
openinference.span.kind	attributes.kind	Span type or category.	String	"llm"
gen_ai.response.model	attributes.model	Model identifier.	String	"gpt-4o"
llm.model_name	attributes.model	Model identifier.	String	"gpt-4o-mini"
ai.model.id	attributes.model	Model identifier.	String	"gpt-4o"
llm.provider	attributes.provider	Model provider name.	String	"openai"
ai.model.provider	attributes.provider	Model provider name.	String	"openai"
gen_ai.request	attributes.model_parameters	Model generation parameters.	Dict	{"temperature":0.7,"max_tokens":256}
llm.invocation_parameters	attributes.model_parameters	Model invocation parameters.	Dict	{"temperature":0.2}
wandb.display_name	display_name	Custom display name for UI.	String	"User Message"
gcp.vertex.agent.session_id	thread_id	Session or thread identifier.	String	"thread_123"
wandb.thread_id	thread_id	Thread identifier for conversations.	String	"thread_123"
wb_run_id	wb_run_id	Associated W&B run identifier.	String	"abc123"
wandb.wb_run_id	wb_run_id	Associated W&B run identifier.	String	"abc123"
gcp.vertex.agent.session_id	is_turn	Marks span as conversation turn.	Boolean	true
wandb.is_turn	is_turn	Marks span as conversation turn.	Boolean	true
langfuse.startTime	start_time (override)	Override span start timestamp.	Timestamp (ISO8601/unix ns)	"2024-01-01T12:00:00Z"
langfuse.endTime	end_time (override)	Override span end timestamp.	Timestamp (ISO8601/unix ns)	"2024-01-01T12:00:01Z"

Limitations

• The Weave UI does not support rendering OTel trace tool calls the Chat view. They appear as raw JSON instead.