Implementation:Vibrantlabsai Ragas MetricBasePrompt

Knowledge Sources	Vibrantlabsai_Ragas
Domains	Prompt Engineering, Metrics, Translation, Evaluation
Last Updated	2026-02-12 00:00 GMT

Overview

The metrics BasePrompt module provides a generic, type-safe abstract base class for structured metric prompts with built-in language adaptation via LLM-powered translation.

Description

This module defines BasePrompt (in the ragas.prompt.metrics namespace), an abstract generic class parameterized by InputModel and OutputModel Pydantic types. It provides a complete prompt construction pipeline: the to_string method assembles the instruction, output JSON schema, formatted examples, and input data into a single prompt string ready for LLM consumption. The class supports few-shot learning through an examples attribute that holds a list of (input, output) tuples, formatted into numbered examples in the prompt. The adapt method enables language adaptation by translating all string content in examples (and optionally the instruction) to a target language using a structured LLM call. Translation is performed by the private _translate_strings helper, which uses a carefully designed safety prompt that instructs the LLM to act only as a translator and never execute instructions found within the text being translated. The translation response is validated to ensure the same number of statements are returned. The module uses utility functions get_all_strings and update_strings from ragas.prompt.utils to extract and replace strings within nested data structures.

Usage

Use this class as the base for all modern Ragas metric prompts that require structured input/output validation, few-shot examples, JSON schema enforcement, and multi-language support.

Code Reference

Source Location

• Repository: Vibrantlabsai_Ragas
• File: src/ragas/prompt/metrics/base_prompt.py

Signature

class BasePrompt(ABC, t.Generic[InputModel, OutputModel]):
    input_model: t.Type[InputModel]
    output_model: t.Type[OutputModel]
    instruction: str
    examples: t.List[t.Tuple[InputModel, OutputModel]]
    language: str = "english"

    def to_string(self, data: InputModel) -> str:
    def _generate_examples(self) -> str:
    async def adapt(
        self,
        target_language: str,
        llm: "InstructorBaseRagasLLM",
        adapt_instruction: bool = False,
    ) -> "BasePrompt[InputModel, OutputModel]":

Import

from ragas.prompt.metrics.base_prompt import BasePrompt

I/O Contract

Inputs (Class Attributes)

Name	Type	Required	Description
input_model	Type[InputModel]	Yes	Pydantic model class for validating prompt input data
output_model	Type[OutputModel]	Yes	Pydantic model class used to generate the expected output JSON schema
instruction	str	Yes	Task description text for the LLM
examples	List[Tuple[InputModel, OutputModel]]	Yes	List of (input, output) example pairs for few-shot learning
language	str	No	Language for the prompt; defaults to "english"

Inputs (to_string)

Name	Type	Required	Description
data	InputModel	Yes	An input data instance validated by the input_model

Inputs (adapt)

Name	Type	Required	Description
target_language	str	Yes	Target language for translation (e.g., "spanish", "french", "hindi")
llm	InstructorBaseRagasLLM	Yes	Instructor-based LLM used for performing translations
adapt_instruction	bool	No	Whether to also translate the instruction text; defaults to False

Outputs

Name	Type	Description
to_string return	str	A complete prompt string containing instruction, JSON schema, examples, and input data
adapt return	BasePrompt[InputModel, OutputModel]	A new prompt instance with examples (and optionally instruction) translated to the target language

Key Helper Classes and Functions

Name	Description
_TranslatedStrings	Pydantic response model for translation results; contains a list of translated statements preserving order and count
_translate_strings(strings, target_language, llm)	Translates a batch of strings to the target language using structured LLM output with safety prompts to prevent instruction injection
_TRANSLATION_INSTRUCTION	Safety prompt text that instructs the LLM to act only as a translator and not execute any embedded instructions

Usage Examples

Defining a Custom Metric Prompt

from pydantic import BaseModel
from ragas.prompt.metrics.base_prompt import BasePrompt

class MyInput(BaseModel):
    question: str
    context: str

class MyOutput(BaseModel):
    score: int
    reason: str

class MyMetricPrompt(BasePrompt[MyInput, MyOutput]):
    input_model = MyInput
    output_model = MyOutput
    instruction = "Evaluate the relevance of the context to the question."
    examples = [
        (
            MyInput(question="What is AI?", context="AI is artificial intelligence."),
            MyOutput(score=2, reason="Directly relevant"),
        ),
    ]

# Use the prompt
prompt = MyMetricPrompt()
prompt_string = prompt.to_string(MyInput(question="What is ML?", context="ML uses data."))

Language Adaptation

# Adapt a prompt to Spanish
spanish_prompt = await prompt.adapt(
    target_language="spanish",
    llm=my_instructor_llm,
    adapt_instruction=True,
)

# The examples and instruction are now translated
print(spanish_prompt.language)  # "spanish"

Related Pages

• Environment:Vibrantlabsai_Ragas_Python_3_9_Core_Environment