Inferences and Schema

Detailed descriptions of classes and methods related to Phoenix inferences and schemas

phoenix.Inferences

class Inferences(
    dataframe: pandas.DataFrame,
    schema: Schema,
    name: Optional[str] = None,
)

A collection of inferences containing a split or cohort of data to be analyzed independently or compared to another cohort. Common examples include training, validation, test, or production data.

[source]

Parameters

• dataframe (pandas.DataFrame): The data to be analyzed or compared.

• schema (Schema): A schema that assigns the columns of the dataframe to the appropriate model dimensions (features, predictions, actuals, etc.).

• name (Optional[str]): The name used to identify the inferences in the application. If not provided, a random name will be generated.

Attributes

• dataframe (pandas.DataFrame): The pandas dataframe of the inferences.

• schema (Schema): The schema of the inferences.

• name (str): The name of the inferences.

The input dataframe and schema are lightly processed during inference initialization and are not necessarily identical to the corresponding dataframe and schema attributes.

Usage

Define inferences ds from a pandas dataframe df and a schema object schema by running

ds = px.Inferences(df, schema)

Alternatively, provide a name for the inferences that will appear in the application:

ds = px.Inferences(df, schema, name="training")

ds is then passed as the primary or reference argument to launch_app.

phoenix.Schema

class Schema(
    prediction_id_column_name: Optional[str] = None,
    timestamp_column_name: Optional[str] = None,
    feature_column_names: Optional[List[str]] = None,
    tag_column_names: Optional[List[str]] = None,
    prediction_label_column_name: Optional[str] = None,
    prediction_score_column_name: Optional[str] = None,
    actual_label_column_name: Optional[str] = None,
    actual_score_column_name: Optional[str] = None,
    prompt_column_names: Optional[EmbeddingColumnNames] = None
    response_column_names: Optional[EmbeddingColumnNames] = None
    embedding_feature_column_names: Optional[Dict[str, EmbeddingColumnNames]] = None,
    excluded_column_names: Optional[List[str]] = None,
)

Assigns the columns of a pandas dataframe to the appropriate model dimensions (predictions, actuals, features, etc.). Each column of the dataframe should appear in the corresponding schema at most once.

[source]

Parameters

• prediction_id_column_name (Optional[str]): The name of the dataframe's prediction ID column, if one exists. Prediction IDs are strings that uniquely identify each record in Phoenix inferences (equivalently, each row in the dataframe). If no prediction ID column name is provided, Phoenix will automatically generate unique UUIDs for each record of the inferences upon Inferences initialization.

• timestamp_column_name (Optional[str]): The name of the dataframe's timestamp column, if one exists. Timestamp columns must be pandas Series with numeric, datetime or object dtypes.

  • If the timestamp column has numeric dtype (int or float), the entries of the column are interpreted as Unix timestamps, i.e., the number of seconds since midnight on January 1st, 1970.

  • If the column has datetime dtype and contains timezone-naive timestamps, Phoenix assumes those timestamps belong to the local timezone and converts them to UTC.

  • If the column has datetime dtype and contains timezone-aware timestamps, those timestamps are converted to UTC.

  • If the column has object dtype having ISO8601 formatted timestamp strings, those entries are converted to datetime dtype UTC timestamps; if timezone-naive then assumed as belonging to local timezone.

  • If no timestamp column is provided, each record in the inferences is assigned the current timestamp upon Inferences initialization.

• feature_column_names (Optional[List[str]]): The names of the dataframe's feature columns, if any exist. If no feature column names are provided, all dataframe column names that are not included elsewhere in the schema and are not explicitly excluded in excluded_column_names are assumed to be features.

• tag_column_names (Optional[List[str]]): The names of the dataframe's tag columns, if any exist. Tags, like features, are attributes that can be used for filtering records of the inferences while using the app. Unlike features, tags are not model inputs and are not used for computing metrics.

• prediction_label_column_name (Optional[str]): The name of the dataframe's predicted label column, if one exists. Predicted labels are used for classification problems with categorical model output.

• prediction_score_column_name (Optional[str]): The name of the dataframe's predicted score column, if one exists. Predicted scores are used for regression problems with continuous numerical model output.

• actual_label_column_name (Optional[str]): The name of the dataframe's actual label column, if one exists. Actual (i.e., ground truth) labels are used for classification problems with categorical model output.

• actual_score_column_name (Optional[str]): The name of the dataframe's actual score column, if one exists. Actual (i.e., ground truth) scores are used for regression problems with continuous numerical output.

• prompt_column_names (Optional[EmbeddingColumnNames]): An instance of EmbeddingColumnNames delineating the column names of an LLM model's prompt embedding vector, prompt text, and optionally links to external resources.

• response_column_names (Optional[EmbeddingColumnNames]): An instance of EmbeddingColumnNames delineating the column names of an LLM model's response embedding vector, response text, and optionally links to external resources.

• embedding_feature_column_names (Optional[Dict[str, EmbeddingColumnNames]]): A dictionary mapping the name of each embedding feature to an instance of EmbeddingColumnNames if any embedding features exist, otherwise, None. Each instance of EmbeddingColumnNames associates one or more dataframe columns containing vector data, image links, or text with the same embedding feature. Note that the keys of the dictionary are user-specified names that appear in the Phoenix UI and do not refer to columns of the dataframe.

• excluded_column_names (Optional[List[str]]): The names of the dataframe columns to be excluded from the implicitly inferred list of feature column names. This field should only be used for implicit feature discovery, i.e., when feature_column_names is unused and the dataframe contains feature columns not explicitly included in the schema.

Usage

See the guide on how to create your own dataset for examples.

phoenix.EmbeddingColumnNames

class EmbeddingColumnNames(
    vector_column_name: str,
    raw_data_column_name: Optional[str] = None,
    link_to_data_column_name: Optional[str] = None,
)

A dataclass that associates one or more columns of a dataframe with an embedding feature. Instances of this class are only used as values in a dictionary passed to the embedding_feature_column_names field of Schema.

[source]

Parameters

• vector_column_name (str): The name of the dataframe column containing the embedding vector data. Each entry in the column must be a list, one-dimensional NumPy array, or pandas Series containing numeric values (floats or ints) and must have equal length to all the other entries in the column.

• raw_data_column_name (Optional[str]): The name of the dataframe column containing the raw text associated with an embedding feature, if such a column exists. This field is used when an embedding feature describes a piece of text, for example, in the context of NLP.

• link_to_data_column_name (Optional[str]): The name of the dataframe column containing links to images associated with an embedding feature, if such a column exists. This field is used when an embedding feature describes an image, for example, in the context of computer vision.

See here for recommendations on handling local image files.

Usage

See the guide on how to create embedding features for examples.

phoenix.TraceDataset

class TraceDataset(
    dataframe: pandas.DataFrame,
    name: Optional[str] = None,
)

Wraps a dataframe that is a flattened representation of spans and traces. Note that it does not require a Schema. See LLM Traces on how to monitor your LLM application using traces. Because Phoenix can also receive traces from your LLM application directly in real time, TraceDataset is mostly used for loading trace data that has been previously saved to file.

[source]

Parameters

• dataframe (pandas.dataframe): a dataframe each row of which is a flattened representation of a span. See LLM Traces for more on traces and spans.

• name (str): The name used to identify the dataset in the application. If not provided, a random name will be generated.

Attributes

• dataframe (pandas.dataframe): a dataframe each row of which is a flattened representation of a span. See LLM Traces for more on traces and spans.

• name (Optional[str]): The name used to identify the dataset in the application.

Usage

The code snippet below shows how to read data from a trace.jsonl file into a TraceDataset, and then pass the dataset to Phoenix through launch_app . Each line of the trace.jsol file is a JSON string representing a span.

from phoenix.trace.utils import json_lines_to_df

with open("trace.jsonl", "r") as f:
    trace_ds = TraceDataset(json_lines_to_df(f.readlines()))
px.launch_app(trace=trace_ds)