OpenInference Ollama Instrumentation

Python auto-instrumentation library for the Ollama Python client.

The traces emitted by this instrumentation are fully OpenTelemetry compatible and can be sent to an OpenTelemetry collector for viewing, such as Arize Phoenix or Arize AX.

What is instrumented

chat calls made through ollama.chat, ollama.Client.chat, and ollama.AsyncClient.chat are exported as OpenInference LLM spans (named Chat and AsyncChat respectively), capturing:

• Input and output messages (llm.input_messages.*, llm.output_messages.*), including tool calls
• Streaming (stream=True): the span finishes when the stream is exhausted, fails, or is abandoned, with the output message and token counts reconstructed from the accumulated chunks
• Tool definitions as llm.tools.N.tool.json_schema — plain Python functions passed via tools=[...] are converted to their JSON schemas
• llm.provider (ollama) and llm.model_name (recorded from the request as well, so errored calls still carry the model)
• Token counts: prompt_eval_count → llm.token_count.prompt, eval_count → llm.token_count.completion, with the total derived when both are present
• llm.invocation_parameters (request options other than messages, model, and tools)
• Errors: exceptions set the span status to ERROR and are recorded as span events

Not currently instrumented: generate, embed/embeddings, and other client methods.

Context attributes (session, user, metadata, tags via using_attributes) propagate onto spans, and sensitive data can be masked with a TraceConfig, e.g. OllamaInstrumentor().instrument(tracer_provider=tracer_provider, config=TraceConfig(hide_inputs=True)). Calls made inside with suppress_tracing(): are not traced.

Installation

pip install openinference-instrumentation-ollama

PyPI package: openinference-instrumentation-ollama

Requires ollama >= 0.4.0.

Quickstart

Install packages needed for this demonstration.

pip install openinference-instrumentation-ollama ollama arize-phoenix opentelemetry-sdk opentelemetry-exporter-otlp

Install and start the Ollama server (pip install ollama installs only the client), then pull a model. The server listens on http://localhost:11434 by default.

ollama pull llama3.2

Start Phoenix in the background as a collector. By default, it listens on http://localhost:6006. (Phoenix does not send data over the internet. It only operates locally on your machine.)

phoenix serve

Set up OllamaInstrumentor to trace your application and send the traces to Phoenix.

from openinference.instrumentation.ollama import OllamaInstrumentor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

endpoint = "http://127.0.0.1:6006/v1/traces"
tracer_provider = TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter(endpoint)))

OllamaInstrumentor().instrument(tracer_provider=tracer_provider)

Run a chat completion against the locally running Ollama server.

import ollama

response = ollama.chat(
    model="llama3.2",
    messages=[{"role": "user", "content": "Why is the sky blue?"}],
)
print(response.message.content)

Now view your traces in the Phoenix UI at http://localhost:6006.

Examples

The examples/ directory contains runnable scripts. They require a running Phoenix and Ollama server, read the model from the OLLAMA_MODEL environment variable (default llama3.2), and send traces to a Phoenix project named ollama-examples.

pip install -r examples/requirements.txt
OLLAMA_MODEL=llama3.2 python examples/chat.py

Development

From the python/ directory: tox run -e test-ollama runs the tests, and tox run -e ruff-mypy-test-ollama runs all checks.

More Info

• More info on OpenInference and Phoenix
• More info on OpenInference and Arize AX
• How to customize spans to track sessions, metadata, etc.
• How to account for private information and span payload customization
• Ollama Python client