How Tracing Works
Your LLM application handles requests — each one might call a model, retrieve documents, run tools, and return a response. A trace captures that entire journey as a tree of spans. Each span represents one operation (an LLM call, a retrieval, a tool invocation) with its input, output, timing, and metadata.- OpenTelemetry (OTel) — the universal, vendor-agnostic framework for collecting and exporting telemetry data. Your instrumentation isn’t locked to Arize — it works with any OTel-compatible backend.
- OpenInference — GenAI-specific semantic conventions built on top of OTel, created by Arize. OpenInference defines the attributes that make traces meaningful for LLM applications: span kinds, message formats, token counts, cost, and more. Arize accepts standard OTel spans and reads OpenInference attributes for richer GenAI visualization.
- Instrumentation wraps your function calls (automatically via integrations, or manually) and captures span data following OpenInference semantic conventions.
- An exporter sends the spans to Arize via OTLP (gRPC by default).
- The Arize collector ingests and visualizes them so you can explore, filter, and debug.
What’s Inside a Trace
Now that you know how data flows from your app to Arize, let’s look at what a trace contains. Each span in the tree is assigned a span kind — the type of operation it represents:| Span Kind | Description |
|---|---|
| LLM | Call to an LLM for a completion or chat |
| Tool | API or function invoked on behalf of an LLM |
| Agent | Root span containing a set of LLM and tool invocations |
| Retriever | Data retrieval query for context from a datastore |
| Chain | The starting point and link between application steps |
| Embedding | Encoding of unstructured data |
| Guardrail | Validates LLM inputs/outputs for safety and compliance |
| Reranker | Relevance-based re-ordering of documents |
| Evaluator | Evaluation process, type, and results |
| Audio | Audio or voice processing operations |
Span Attributes
Each span also carries attributes — key-value pairs likellm.model_name, llm.input_messages, input.value, output.value. These are the OpenInference semantic conventions — auto-instrumentation sets them for you, and you can add your own via Customize your traces. Check the OpenInference source for every attribute defined per span kind: