> ## Documentation Index
> Fetch the complete documentation index at: https://docs.noxus.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Observability Providers

> Connect tracing backends to get full visibility into your AI model calls

## What are Observability Providers?

Observability providers connect Noxus to external tracing backends so you can inspect every LLM call made by your flows and agents — inputs, outputs, token usage, latency, fallback routing, and human feedback scores — all in one place.

<Info>
  Noxus instruments all AI activity using OpenTelemetry and the OpenInference
  semantic conventions. Any backend that speaks OTLP receives rich, structured
  traces without additional configuration.
</Info>

Each provider you connect receives a full copy of every trace. You can connect multiple backends simultaneously — useful for routing to both an internal Phoenix instance and a team Langfuse project at the same time.

## Supported Backends

<Tabs>
  <Tab title="Arize Phoenix">
    **Open-source LLM observability** — ideal for self-hosted deployments or teams that want full control over their trace data.

    * Self-hostable or available as [Arize Phoenix Cloud](https://phoenix.arize.com)
    * Rich UI for trace inspection, span timelines, and prompt analysis
    * Supports human feedback annotations directly from the Noxus chat interface
    * Authentication via Bearer token (optional for local instances)
  </Tab>

  <Tab title="Langfuse">
    **Open-source LLM engineering platform** — includes tracing, prompt management, and evaluation workflows.

    * Self-hostable or available as [Langfuse Cloud](https://langfuse.com)
    * Supports feedback scoring from the Noxus chat interface
    * Authentication via public key + secret key pair
    * Endpoint should point to your Langfuse instance's OTLP ingest path (e.g. `https://cloud.langfuse.com/api/public/otel/v1/traces`)
  </Tab>

  <Tab title="Generic OpenTelemetry">
    **Any OTLP-compatible backend** — connect Grafana Tempo, Honeycomb, Jaeger, Datadog, or any other collector that accepts OTLP over HTTP.

    * Provide the full OTLP endpoint URL (e.g. `https://api.honeycomb.io/v1/traces`)
    * Optional Bearer token for authentication
    * Project or service name used as the OTel resource attribute
  </Tab>
</Tabs>

## Connecting a Provider

<Steps>
  <Step title="Open provider settings">
    Navigate to **Organization Settings** > **Providers** and click **Add Provider**. Select the observability category.
  </Step>

  <Step title="Choose your backend">
    Pick **Phoenix**, **Langfuse**, or **OpenTelemetry** from the provider list.
  </Step>

  <Step title="Enter connection details">
    Fill in the endpoint URL, project name, and any required credentials. Noxus tests the connection by sending a probe span before saving.
  </Step>

  <Step title="Activate">
    Once the connection test passes, the provider is marked active and tracing begins immediately for all new flow and agent executions.
  </Step>
</Steps>

<Note>
  The endpoint must be the full OTLP path — Noxus does not append `/v1/traces` automatically. Double-check your backend's documentation for the exact ingest URL.
</Note>

## What Gets Traced

Every AI execution in Noxus produces a structured trace with the following spans:

| Span                 | Description                                                                                |
| :------------------- | :----------------------------------------------------------------------------------------- |
| **Agent**            | Top-level span for the entire node execution. Captures the full prompt and final response. |
| **LLM call**         | One span per model invocation, including the model ID used, token counts, and latency.     |
| **Fallback routing** | Recorded when the platform falls back from one model to another within a preset.           |
| **Cache hit**        | Recorded when a cached response is served, with the cache key and saved latency.           |
| **Tool call**        | Spans for each tool or function call made during an agent turn.                            |

All spans follow the [OpenInference](https://github.com/Arize-ai/openinference) semantic conventions, making them compatible with any OpenInference-aware UI.

## Human Feedback Scoring

For **Phoenix** and **Langfuse** providers, users can submit thumbs-up / thumbs-down feedback on agent responses directly from the Noxus chat interface. Feedback is sent to the backend as a span annotation or score linked to the originating trace, letting you filter and analyze low-quality responses in your observability tool.

## Multiple Providers

You can connect more than one observability provider at the same time. Noxus fans out every trace to all active backends in parallel — each backend receives an identical copy of the span data. This is useful for:

* Sending traces to both a self-hosted Phoenix instance and a team Langfuse project
* Running a new backend alongside an existing one during evaluation
* Separating production and staging observability into different projects on the same backend

<CardGroup cols={2}>
  <Card title="Models & Providers" icon="server" href="/core/infrastructure/models-providers">
    Configure the AI model providers that power your flows.
  </Card>

  <Card title="Operations" icon="chart-line" href="/core/platform/operations">
    Platform-level monitoring, audit logs, and scaling.
  </Card>
</CardGroup>