DSPy

Instrument and observe your DSPy application via the DSPyInstrumentor

DSPy is a framework for automatically prompting and fine-tuning language models. It provides composable and declarative APIs that allow developers to describe the architecture of their LLM application in the form of a "module" (inspired by PyTorch's nn.Module). It them compiles these modules using "teleprompters" that optimize the module for a particular task. The term "teleprompter" is meant to evoke "prompting at a distance," and could involve selecting few-shot examples, generating prompts, or fine-tuning language models.

Phoenix makes your DSPy applications observable by visualizing the underlying structure of each call to your compiled DSPy module.

Launch Phoenix

Phoenix Cloud

Sign up for Phoenix:

Sign up for an Arize Phoenix account at https://app.phoenix.arize.com/login

Install packages:

pip install arize-phoenix-otel

Set your Phoenix endpoint and API Key:

import os

# Add Phoenix API Key for tracing
PHOENIX_API_KEY = "ADD YOUR API KEY"
os.environ["PHOENIX_CLIENT_HEADERS"] = f"api_key={PHOENIX_API_KEY}"
os.environ["PHOENIX_COLLECTOR_ENDPOINT"] = "https://app.phoenix.arize.com"

Your Phoenix API key can be found on the Keys section of your dashboard.

Command Line

Launch your local Phoenix instance:

pip install arize-phoenix
phoenix serve

For details on customizing a local terminal deployment, see Terminal Setup.

Install packages:

pip install arize-phoenix-otel

Set your Phoenix endpoint:

import os

os.environ["PHOENIX_COLLECTOR_ENDPOINT"] = "http://localhost:6006"

See Terminal for more details

Docker

Pull latest Phoenix image from Docker Hub:

docker pull arizephoenix/phoenix:latest

Run your containerized instance:

docker run -p 6006:6006 arizephoenix/phoenix:latest

This will expose the Phoenix on localhost:6006

Install packages:

pip install arize-phoenix-otel

Set your Phoenix endpoint:

import os

os.environ["PHOENIX_COLLECTOR_ENDPOINT"] = "http://localhost:6006"

For more info on using Phoenix with Docker, see Docker.

Notebook

Install packages:

pip install arize-phoenix

Launch Phoenix:

import phoenix as px
px.launch_app()

By default, notebook instances do not have persistent storage, so your traces will disappear after the notebook is closed. See self-hosting or use one of the other deployment options to retain traces.

Install

pip install openinference-instrumentation-dspy openinference-instrumentation-litellm dspy

DSPy uses LiteLLM under the hood to make some calls. By adding the OpenInference library for LiteLLM, you'll be able to see additional information like token counts on your traces.

Setup

Connect to your Phoenix instance using the register function.

from phoenix.otel import register

# configure the Phoenix tracer
tracer_provider = register(
  project_name="my-llm-app", # Default is 'default'
  auto_instrument=True # Auto-instrument your app based on installed OI dependencies
)

Run DSPy

Now run invoke your compiled DSPy module. Your traces should appear inside of Phoenix.

class BasicQA(dspy.Signature):
    """Answer questions with short factoid answers."""

    question = dspy.InputField()
    answer = dspy.OutputField(desc="often between 1 and 5 words")


if __name__ == "__main__":
    turbo = dspy.OpenAI(model="gpt-3.5-turbo")

    dspy.settings.configure(lm=turbo)

    with using_attributes(
        session_id="my-test-session",
        user_id="my-test-user",
        metadata={
            "test-int": 1,
            "test-str": "string",
            "test-list": [1, 2, 3],
            "test-dict": {
                "key-1": "val-1",
                "key-2": "val-2",
            },
        },
        tags=["tag-1", "tag-2"],
        prompt_template_version="v1.0",
        prompt_template_variables={
            "city": "Johannesburg",
            "date": "July 11th",
        },
    ):
        # Define the predictor.
        generate_answer = dspy.Predict(BasicQA)

        # Call the predictor on a particular input.
        pred = generate_answer(
            question="What is the capital of the united states?"  # noqa: E501
        )  # noqa: E501
        print(f"Predicted Answer: {pred.answer}")

Observe

Now that you have tracing setup, all predictions will be streamed to your running Phoenix for observability and evaluation.

Traces and spans from an instrumented DSPy custom module.

Resources

• Example notebook
• OpenInference package
• Working examples