Pydantic AI
Send Pydantic AI traces to LiteLLM Lens using the runnable examples in this repository.
Prerequisites
You need a LiteLLM gateway with tracing enabled, a LiteLLM key, and a configured model alias. The swarm example needs a model that supports tool calls. A Lens worker is required for investigations; viewing traces does not require one.
Install uv. It uses the checked-in Python version and resolves each example’s dependencies from its uv workspace.
Configuration
For a fresh checkout:
git clone https://github.com/BerriAI/litellm-lens-example.git
cd litellm-lens-example/pydantic-ai
cp .env.example .env
If you already cloned the repository, run the remaining commands from pydantic-ai/. Copy .env.example to .env if it does not exist, then set:
| Variable | Value |
|---|---|
LITELLM_GATEWAY_URL | Your gateway’s base URL without a trailing slash or /v1, for example http://localhost:4002 |
LITELLM_API_KEY | Your LiteLLM key |
LITELLM_MODEL | A model alias configured on your gateway |
The checked-in values target a local development gateway. Replace them for your deployment. Keep the exporter settings from .env.example; the examples configure their trace exporters in code. They send traces to LITELLM_GATEWAY_URL/v1/traces with the LiteLLM key as a bearer token.
Leave MOCK_LITELLM_GATEWAY_URL unset unless you intend to send an additional trace copy to the local recorder.
Run an example
Simple agent
A research_agent answers one question.
uv run --env-file .env --package lens-pydantic-ai-simple simple/main.py
See simple/main.py for the implementation.
Agent swarm
A coordinator delegates to search_agent and writer_agent through tools.
uv run --env-file .env --package lens-pydantic-ai-swarm swarm/main.py
See swarm/main.py for the implementation.
Streaming
Set LITELLM_STREAM=1 to enable streaming in either example:
LITELLM_STREAM=1 uv run --env-file .env --package lens-pydantic-ai-simple simple/main.py
Verify the trace
After the example prints its answer, open Lens > Traces on your gateway and select the new run. Look for the run associated with research_agent. Inspect the input, output, and model spans. For the swarm, inspect the specialist activity described above; its exact span layout depends on the framework.
How tracing works
Pydantic AI emits agent, tool, and model spans through Agent.instrument_all(). The shared gateway transport records request attempts and gateway call IDs for matching model calls to spend.
See the shared gateway transport for request-attempt and spend-correlation details.
Troubleshooting
If model calls fail, check the gateway URL, key, and model alias. If an answer appears but the trace is missing, check the terminal for exporter errors and confirm tracing is enabled on the same gateway. A model call succeeding does not confirm that its trace export succeeded.
