Arize Phoenix: 快速上手指南

brew install uv

pipx install uv

$ uv init my-project
Initialized project `my-project` at `/Users/james/tmp/my-project`

$ tree my-project/
my-project/
├── README.md
├── hello.py
└── pyproject.toml

1 directory, 3 files

$ cd my-project
$ uv add arize-phoenix litellm arize-phoenix-otel openinference-instrumentation-litellm

$ uv run python

import os
from getpass import getpass

os.environ["OPENROUTER_API_KEY"] = getpass("🔑 Enter your OpenRouter API key: ")

import pandas as pd
from phoenix.evals import LiteLLMModel, QAEvaluator, run_evals

df = pd.DataFrame(
    [
        {
            "input": "Where is Oriental Pearl Tower located?",
            "reference": "The Oriental Pearl Tower[a] is a futurist TV tower in Lujiazui, Shanghai.",
            "output": "It's located in Lujiazui, Shanghai.",
        },
        {
            "input": "Is Oriental Pearl Tower the tallest in China?",
            "reference": "Built from 1991 to 1994, the tower was the tallest structure in China until the completion of nearby World Financial Center in 2007.",
            "output": "It's the tallest in China.",
        }
    ]
)

eval_model = LiteLLMModel(model="openrouter/openai/gpt-4o-mini")
qa_evaluator = QAEvaluator(eval_model)

qa_eval_df = run_evals(
    dataframe=df,
    evaluators=[qa_evaluator],
    provide_explanation=True,
)

>>> print(qa_eval_df)
[       label  score                                        explanation
0    correct      1  To determine if the answer is correct, we firs...
1  incorrect      0  To determine whether the answer correctly addr...]

>>> qa_eval_df[0].loc[0, 'explanation']
'To determine if the answer is correct, we first need to analyze the question, the reference text, and the answer provided.\n\n1. **Identify the Question**: The question asks, "Where is Oriental Pearl Tower located?" This is a straightforward inquiry about the geographical location of the Oriental Pearl Tower.\n\n2. **Examine the Reference Text**: The reference text states, "The Oriental Pearl Tower is a futurist TV tower in Lujiazui, Shanghai." This sentence provides a clear location for the Oriental Pearl Tower, specifying that it is in Lujiazui, which is a district in Shanghai.\n\n3. **Analyze the Answer**: The answer given is, "It\'s located in Lujiazui, Shanghai." This response directly addresses the question by repeating the location mentioned in the reference text.\n\n4. **Compare the Answer to the Reference**: The answer matches the information provided in the reference text. It specifies both the district (Lujiazui) and the city (Shanghai), which fully answers the question about the location of the Oriental Pearl Tower.\n\n5. **Conclusion**: Since the answer accurately reflects the information found in the reference text and fully answers the question, we can conclude that the answer is correct.\n\nLABEL: correct'

>>> qa_eval_df[0].loc[1, 'explanation']
'To determine whether the answer correctly addresses the question, we need to analyze both the question and the reference text.\n\n1. **Understanding the Question**: The question asks if the Oriental Pearl Tower is the tallest structure in China. This requires a definitive answer regarding its height in comparison to other structures in China.\n\n2. **Analyzing the Reference Text**: The reference states that the Oriental Pearl Tower was the tallest structure in China until the completion of the World Financial Center in 2007. This indicates that the Oriental Pearl Tower is no longer the tallest structure in China, as it was surpassed by another building.\n\n3. **Evaluating the Answer**: The answer provided is "It\'s the tallest in China." This statement directly contradicts the information in the reference text, which clearly states that the Oriental Pearl Tower was the tallest until 2007, implying that it is not currently the tallest.\n\n4. **Conclusion**: Since the answer claims that the Oriental Pearl Tower is still the tallest in China, but the reference text indicates that it is not, the answer does not correctly answer the question.\n\nLABEL: incorrect'

$ uv run phoenix serve

...omitted...

INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:6006 (Press CTRL+C to quit)

from phoenix.otel import register
from openinference.instrumentation.litellm import LiteLLMInstrumentor

tracer_provider = register(
  project_name="my-project",
  endpoint="http://localhost:6006/v1/traces",
)

LiteLLMInstrumentor().instrument(tracer_provider=tracer_provider)