AWS Bedrock AI Observability installation

Let AI instrument your LLM calls for you

Skip the manual setup — run this in your project and the wizard installs the SDK and wires up AI Observability for you.

Install dependencies

Required

Install the OpenTelemetry SDK, OTLP exporter, and the AWS SDK instrumentation for your language.

pip install boto3 opentelemetry-instrumentation-botocore opentelemetry-sdk "posthog[otel]"
npm install @aws-sdk/client-bedrock-runtime @opentelemetry/instrumentation-aws-sdk @opentelemetry/sdk-node @opentelemetry/resources @posthog/ai

Set up the OpenTelemetry exporter

Required

Configure the OpenTelemetry SDK to export traces to PostHog's OTLP ingestion endpoint. PostHog converts gen_ai.* spans into $ai_generation events automatically.

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.resources import Resource, SERVICE_NAME
from posthog.ai.otel import PostHogSpanProcessor
from opentelemetry.instrumentation.botocore import BotocoreInstrumentor

resource = Resource(attributes={
    SERVICE_NAME: "my-ai-app",
})

provider = TracerProvider(resource=resource)
provider.add_span_processor(
    PostHogSpanProcessor(
        api_key="<ph_project_token>",
        host="https://us.i.posthog.com",
    )
)
trace.set_tracer_provider(provider)

BotocoreInstrumentor().instrument()
import { NodeSDK } from '@opentelemetry/sdk-node'
import { resourceFromAttributes } from '@opentelemetry/resources'
import { PostHogSpanProcessor } from '@posthog/ai/otel'
import { AwsInstrumentation } from '@opentelemetry/instrumentation-aws-sdk'

const sdk = new NodeSDK({
  resource: resourceFromAttributes({
    'service.name': 'my-ai-app',
  }),
  spanProcessors: [
    new PostHogSpanProcessor({
      projectToken: '<ph_project_token>',
      host: 'https://us.i.posthog.com',
    }),
  ],
  instrumentations: [new AwsInstrumentation()],
})
sdk.start()

Call Bedrock

Required

Make Bedrock API calls as normal. The instrumentation automatically captures gen_ai.* spans for Converse, ConverseStream, InvokeModel, and InvokeModelWithResponseStream operations.

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")

response = client.converse(
    modelId="openai.gpt-oss-20b-1:0",
    messages=[
        {
            "role": "user",
            "content": [{"text": "Tell me a fun fact about hedgehogs."}],
        }
    ],
)

for block in response["output"]["message"]["content"]:
    if "text" in block:
        print(block["text"])
        break
// The AWS SDK must be imported after sdk.start() so the
// instrumentation can patch it.
const {
  BedrockRuntimeClient,
  ConverseCommand,
} = await import('@aws-sdk/client-bedrock-runtime')

const client = new BedrockRuntimeClient({ region: 'us-east-1' })

const response = await client.send(
  new ConverseCommand({
    modelId: 'openai.gpt-oss-20b-1:0',
    messages: [
      {
        role: 'user',
        content: [{ text: 'Tell me a fun fact about hedgehogs.' }],
      },
    ],
  })
)

const textBlock = response.output?.message?.content?.find((b) => 'text' in b)
console.log(textBlock?.text)

await sdk.shutdown()

Supported models

The instrumentation emits gen_ai.* spans for Amazon Titan, Amazon Nova, and Anthropic Claude models. Tool call instrumentation is available for Amazon Nova and Anthropic Claude 3+.

Note: If you want to capture LLM events anonymously, omit the posthog_distinct_id. See our docs on anonymous vs identified events to learn more.

You can expect captured $ai_generation events to have the following properties:

Property	Description
`$ai_model`	The specific model, like `gpt-5-mini` or `claude-4-sonnet`
`$ai_latency`	The latency of the LLM call in seconds
`$ai_time_to_first_token`	Time to first token in seconds (streaming only)
`$ai_tools`	Tools and functions available to the LLM
`$ai_input`	List of messages sent to the LLM
`$ai_input_tokens`	The number of tokens in the input (often found in response.usage)
`$ai_output_choices`	List of response choices from the LLM
`$ai_output_tokens`	The number of tokens in the output (often found in `response.usage`)
`$ai_total_cost_usd`	The total cost in USD (input + output)
[...]	See full list of properties

Group traces into sessions

Optional

PostHog groups traces into a session when they share an $ai_session_id. Set it if your product has multi-turn conversations, so the Sessions tab can reconstruct them. Workloads that finish in a single trace, like batch jobs or one-shot generation, do not need it.

The instrumentation creates the LLM span for you, so there is no call to pass the session ID to. Add a span processor that sets the $ai_session_id attribute as each span starts. PostHog forwards span attributes it does not recognize onto the event, so the value arrives as the $ai_session_id property.

import contextvars
from collections.abc import Iterator
from contextlib import contextmanager
from typing import Optional

from opentelemetry.context import Context
from opentelemetry.sdk.trace import Span, SpanProcessor

session_id_var: contextvars.ContextVar[Optional[str]] = contextvars.ContextVar(
    "ai_session_id", default=None
)


class SessionIdSpanProcessor(SpanProcessor):
    def on_start(self, span: Span, parent_context: Optional[Context] = None) -> None:
        session_id = session_id_var.get()
        if session_id is not None:
            span.set_attribute("$ai_session_id", session_id)


@contextmanager
def ai_session(session_id: str) -> Iterator[None]:
    token = session_id_var.set(session_id)
    try:
        yield
    finally:
        session_id_var.reset(token)


# Register it on the same provider as PostHogSpanProcessor
provider.add_span_processor(SessionIdSpanProcessor())

# Resetting on exit keeps the ID off the next request that reuses this thread
with ai_session("conversation-abc"):
    reply = handle_turn(user_message)
import { AsyncLocalStorage } from 'node:async_hooks'
import type { Span, SpanProcessor } from '@opentelemetry/sdk-trace-base'

const sessionStore = new AsyncLocalStorage<string>()

class SessionIdSpanProcessor implements SpanProcessor {
  onStart(span: Span): void {
    const sessionId = sessionStore.getStore()
    if (sessionId) {
      span.setAttribute('$ai_session_id', sessionId)
    }
  }
  onEnd(): void {}
  async shutdown(): Promise<void> {}
  async forceFlush(): Promise<void> {}
}

// Every span started inside the callback carries this session ID
const reply = await sessionStore.run('conversation-abc', () => handleTurn(userMessage))

On Node, add new SessionIdSpanProcessor() to the spanProcessors array of the NodeSDK you configured earlier, next to PostHogSpanProcessor. Keep the rest of that setup as it is, including instrumentations and the sdk.start() call.

If a process only ever handles one session, set $ai_session_id as a resource attribute next to service.name instead. Resource attributes apply to every span the process emits, so that only works when the process and the session are the same thing.

Verify traces and generations
Recommended
Confirm LLM events are being sent to PostHog
Let's make sure LLM events are being captured and sent to PostHog. Under AI Observability, you should see rows of data appear in the Traces and Generations tabs.

Check for LLM events in PostHog

Next steps

Recommended

Now that you're capturing AI conversations, continue with the resources below to learn what else AI Observability enables within the PostHog platform.

Resource	Description
Basics	Learn the basics of how LLM calls become events in PostHog.
Generations	Read about the `$ai_generation` event and its properties.
Traces	Explore the trace hierarchy and how to use it to debug LLM calls.
Spans	Review spans and their role in representing individual operations.
Anaylze LLM performance	Learn how to create dashboards to analyze LLM performance.

AWS Bedrock AI Observability installation

Contents

Install dependencies

Set up the OpenTelemetry exporter

Call Bedrock

Group traces into sessions

Verify traces and generations

Next steps

Still have questions?

Was this page useful?