Implementation:Explodinggradients Ragas DemonstrationConfig And InstructionConfig

Knowledge Sources	Explodinggradients_Ragas
Domains	Configuration, Optimization
Last Updated	2026-02-10 00:00 GMT

Overview

Pydantic configuration models for controlling prompt optimization parameters: demonstration selection and instruction tuning.

Description

DemonstrationConfig controls few-shot example selection during prompt optimization, supporting random or similarity-based techniques with configurable top-k and threshold. InstructionConfig configures instruction optimization with an LLM, loss function, and optimizer (defaults to GeneticOptimizer).

Usage

Use these configuration models when setting up prompt optimization workflows with Ragas optimizers.

Code Reference

Source Location

• Repository: Explodinggradients_Ragas
• File: src/ragas/config.py
• Lines: 15-37

Signature

class DemonstrationConfig(BaseModel):
    embedding: Any  # must be BaseRagasEmbeddings
    enabled: bool = True
    top_k: int = 3
    threshold: float = 0.7
    technique: Literal["random", "similarity"] = "similarity"

class InstructionConfig(BaseModel):
    llm: BaseRagasLLM
    enabled: bool = True
    loss: Optional[Loss] = None
    optimizer: Optimizer = GeneticOptimizer()
    optimizer_config: Dict[str, Any]  # default: {"max_steps": 100}

Import

from ragas.config import DemonstrationConfig, InstructionConfig

I/O Contract

Inputs

Name	Type	Required	Description
embedding	BaseRagasEmbeddings	Yes (DemonstrationConfig)	Embedding model for similarity-based selection
llm	BaseRagasLLM	Yes (InstructionConfig)	LLM for instruction optimization
loss	Optional[Loss]	No	Loss function for optimization
optimizer	Optimizer	No	Optimizer instance (default: GeneticOptimizer)

Outputs

Name	Type	Description
DemonstrationConfig	BaseModel	Configuration for demonstration/few-shot selection
InstructionConfig	BaseModel	Configuration for instruction optimization

Usage Examples

from ragas.config import DemonstrationConfig, InstructionConfig
from ragas.llms import llm_factory
from ragas.embeddings import embedding_factory

llm = llm_factory()
embeddings = embedding_factory()

demo_config = DemonstrationConfig(
    embedding=embeddings,
    technique="similarity",
    top_k=5,
    threshold=0.8,
)

instruction_config = InstructionConfig(
    llm=llm,
    optimizer_config={"max_steps": 50},
)

Related Pages

• Principle:Explodinggradients_Ragas_Genetic_Prompt_Optimization