Architecture

AIO uses an OpenTelemetry-based pipeline to move telemetry from your AI application into Acceldata for storage and analysis.

Architecture summary

In short:

The SDK captures → the collector receives → ClickHouse stores → AIO presents.

This is probably the right level for the Architecture page. The Introduction explains what traces and spans are and how to instrument an application; Architecture explains how that telemetry moves through the system.

At a high level, traces flow through the following components:

How traces flow

AIO SDK

The AIO SDK instruments your AI application and creates or enriches OpenTelemetry spans for supported operations.

These spans can represent activity such as model calls, tool calls, MCP interactions, and outbound HTTP requests.

The SDK then exports the telemetry to an OpenTelemetry Collector.

For information about installing and configuring the SDK, see Introduction.

OpenTelemetry Collector

The OpenTelemetry Collector receives trace data exported by the SDK.

It acts as the ingestion layer between your application and AIO storage.

The collector writes the received trace data to ClickHouse.

ClickHouse

ClickHouse is the primary storage engine for AIO trace data.

It stores the traces and spans ingested through the OpenTelemetry Collector.

The AIO service queries this stored data when you view observability information in the product.

AIO Service and UI

The AIO Service queries trace data stored in ClickHouse and makes it available through the AIO user interface.

From the UI, you can inspect information such as:

• traces
• spans
• sessions
• model and tool activity
• token information
• Errors

Deployment approaches

Both deployment offerings share the identical trace path architecture:

• Acceldata Hosted: Fully managed solution hosted and maintained by Acceldata.

• Bring Your Own Data Plane: Deployment option allowing customers to keep the data plane in their own infrastructure while utilizing the AIO service and UI.