Implementation:Arize ai Phoenix Legacy Templates

Overview

The Legacy Templates module implements the prompt templating system for the Phoenix Evals subsystem. It provides PromptTemplate for basic prompt construction with variable substitution, ClassificationTemplate (extending PromptTemplate) for classification tasks with rails, scores, and explanation support, and MultimodalPrompt for representing prompts that mix text, audio, and image content.

The module supports configurable delimiters for variable placeholders (default curly braces), dot-key formatting via a custom DotKeyFormatter (allowing keys with dots as single identifiers), and chain-of-thought label parsing via parse_label_from_chain_of_thought_response(). Helper functions normalize_classification_template() and normalize_prompt_template() ensure consistent template types, and map_template() renders a template across all rows of a DataFrame.

Supporting data classes include PromptOptions (controlling explanation mode), PromptPartContentType (enum for text/audio/image), PromptPart (a rendered prompt segment), and PromptPartTemplate (a template segment with content type).

Code Reference

Attribute	Details
Source File	`packages/phoenix-evals/src/phoenix/evals/legacy/templates.py`
Repository	Arize-ai/phoenix
Lines	289
Module	`phoenix.evals.legacy.templates`
Key Symbols	`PromptTemplate`, `ClassificationTemplate`, `MultimodalPrompt`, `PromptOptions`, `PromptPartContentType`, `DotKeyFormatter`, `parse_label_from_chain_of_thought_response()`, `normalize_classification_template()`, `normalize_prompt_template()`, `map_template()`
Dependencies	`pandas`, `phoenix.evals.legacy.exceptions`

I/O Contract

PromptTemplate

Method	Parameters	Returns
`__init__(template, delimiters, variables)`	`template: Union[str, List[PromptPartTemplate]]`, `delimiters: Tuple[str, str]` (default `("{", "}")`), `variables: Optional[List[str]]`	None
`prompt(options)`	`options: Optional[PromptOptions]`	`List[PromptPartTemplate]` - the raw template parts
`format(variable_values, options)`	`variable_values: Mapping[str, Union[bool, int, float, str]]`, `options: Optional[PromptOptions]`	`MultimodalPrompt` - rendered prompt with variables substituted

ClassificationTemplate

Method	Parameters	Returns
`__init__(rails, template, explanation_template, explanation_label_parser, delimiters, scores)`	`rails: List[str]`, `template: Union[str, List[PromptPartTemplate]]`, `explanation_template: Optional[...]`, `explanation_label_parser: Optional[Callable]`, `delimiters: Tuple[str, str]`, `scores: Optional[List[float]]`	None
`prompt(options)`	`options: Optional[PromptOptions]`	`List[PromptPartTemplate]` - returns explanation template when `provide_explanation=True`
`extract_label_from_explanation(raw_string)`	`raw_string: str`	`str` - extracted label from chain-of-thought response
`score(rail)`	`rail: str`	`float` - score for the given rail (0.0 if not found)

MultimodalPrompt

Method	Parameters	Returns
`from_string(string_prompt)`	`string_prompt: str`	`MultimodalPrompt` - wraps string as a text-only prompt
`to_text_only_prompt()`	None	`str` - concatenated text parts; raises `ValueError` if non-text parts exist
`__str__()`	None	`str` - all parts concatenated

PromptPartContentType Enum

Value	Description
`TEXT`	Plain text content
`AUDIO`	Audio content (base64 encoded)
`IMAGE`	Image content (base64 encoded)

parse_label_from_chain_of_thought_response()

Parameter	Type	Description
`raw_string`	`str`	Raw LLM response containing "LABEL:" followed by the classification.
Returns	`str`	Extracted label text, or the full string if no "LABEL" keyword found.

The function searches for the keyword "LABEL" (case-insensitive), strips leading punctuation, and returns the text following it up to the next newline or "EXPLANATION" keyword (handling cases where the model produces label before or after explanation).

DotKeyFormatter

A custom string.Formatter subclass that treats the entire field name (including dots) as a single dictionary key, preventing Python's default attribute/item access splitting on dots.

Usage Examples

from phoenix.evals.legacy.templates import (
    PromptTemplate,
    ClassificationTemplate,
    MultimodalPrompt,
    PromptOptions,
    parse_label_from_chain_of_thought_response,
)

# Basic prompt template
template = PromptTemplate(template="Evaluate this text: {input}")
prompt = template.format(variable_values={"input": "Hello world"})
print(str(prompt))  # "Evaluate this text: Hello world"

# Classification template with explanation support
cls_template = ClassificationTemplate(
    rails=["positive", "negative"],
    template="Classify sentiment: {text}",
    explanation_template="Explain then classify sentiment: {text}\nEXPLANATION:",
    scores=[1, 0],
)

# Get base prompt
base = cls_template.prompt(PromptOptions(provide_explanation=False))

# Get explanation prompt
explanation = cls_template.prompt(PromptOptions(provide_explanation=True))

# Parse chain-of-thought response
label = parse_label_from_chain_of_thought_response(
    "EXPLANATION: The text is clearly happy.\nLABEL: positive"
)
# label = "positive"

# Multimodal prompt
from phoenix.evals.legacy.templates import (
    PromptPartTemplate,
    PromptPartContentType,
)

multimodal_template = PromptTemplate(
    template=[
        PromptPartTemplate(
            content_type=PromptPartContentType.TEXT,
            template="Describe this image: ",
        ),
        PromptPartTemplate(
            content_type=PromptPartContentType.IMAGE,
            template="{image_data}",
        ),
    ]
)

Related Pages

Arize_ai_Phoenix_Legacy_Default_Templates - Predefined ClassificationTemplate instances
Arize_ai_Phoenix_Legacy_Classify - Primary consumer via llm_classify()
Arize_ai_Phoenix_Legacy_Generate - Uses PromptTemplate for generation tasks
Arize_ai_Phoenix_Legacy_Span_Templates - Span-level ClassificationTemplate instances
Arize_ai_Phoenix_Legacy_Audio_Templates - Multimodal audio ClassificationTemplate

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment