🪆Nested Configurations

Hypster enables hierarchical configuration management through the hp.nest() method, allowing you to compose complex configurations from smaller, reusable components.

For an in depth tutorial, please check out the article on Medium: Implementing Modular-RAG using Haystack and Hypster

nest Function Signature

def nest(
    config_func: Union[str, Path, "Hypster"],
    *,
    name: Optional[str] = None,
    final_vars: List[str] = [],
    exclude_vars: List[str] = [],
    values: Dict[str, Any] = {}
) -> Dict[str, Any]

Parameters

• config_func: Either a path to a saved configuration or a Hypster config object

• name: Optional name for the nested configuration (used in dot notation)

• final_vars: List of variables that cannot be modified by parent configs

• exclude_vars: List of variables to exclude from the configuration

• values: Dictionary of values to override in the nested configuration

Steps for nesting

1

Define a reusable config

from hypster import config, HP

@config
def llm_config(hp: HP):
    model = hp.select({
        "haiku": "claude-3-haiku-20240307",
        "sonnet": "claude-3-sonnet-20240229"
    }, default="haiku")
    temperature = hp.number(0.7, min=0, max=1)

2

Save it

llm_config.save("configs/llm.py")

3

Define a parent config and use hp.nest

@config
def qa_config(hp: HP):
    # Load and nest LLM configuration
    llm = hp.nest("configs/llm.py")

    # Add QA-specific parameters
    max_context_length = hp.int(1000, min=100, max=2000)

    # Combine LLM and QA parameters
    qa_pipeline = QAPipeline(
        model=llm["model"],
        temperature=llm["temperature"],
        max_context_length=max_context_length
    )

4

Instantiate using dot notation

qa_config(values={
    "llm.model": "sonnet",
    "llm.temperature": 0.5,
    "max_context_length": 1500
})

Configuration Sources

hp.nest() accepts two types of sources:

Path to Configuration File

llm = hp.nest("configs/llm.py")

Direct Configuration Object

from hypster import load

# Load the configuration
llm_config = load("configs/llm.py")

# Use the loaded config
qa_config = hp.nest(llm_config)

Value Assignment

Values for nested configurations can be set using either dot notation or nested dictionaries:

# Using dot notation
qa_config(values={
    "llm.model": "sonnet",
    "llm.temperature": 0.5,
    "max_context_length": 1500
})

# Using nested dictionary
qa_config(values={
    "llm": {
        "model": "sonnet",
        "temperature": 0.5
    },
    "max_context_length": 1500
})

Hierarchical Nesting

Configurations can be nested multiple times to create modular, reusable components:

@config
def indexing_config(hp: HP):
    # Reuse LLM config for document processing
    llm = hp.nest("configs/llm.py")

    # Indexing-specific parameters
    embedding_dim = hp.int(512, min=128, max=1024)

    # Process documents with LLM
    enriched_docs = process_documents(
        llm=llm["model"],
        temperature=llm["temperature"],
        embedding_dim=embedding_dim
    )

@config
def rag_config(hp: HP):
    # Reuse indexing config (which includes LLM config)
    indexing = hp.nest("configs/indexing.py")

    # Add retrieval configuration
    retrieval = hp.nest("configs/retrieval.py")

Passing Values to Nested Configs

Use the values parameter to pass dependent values to nested configuration values:

retrieval = hp.nest(
    "configs/retrieval.py",
    values={
        "embedding_dim": indexing["embedding_dim"],
        "top_k": 5
    }
)

final_vars and exclude_vars are also supported.

Best Practices

1. Modular Design

   • Create small, focused configurations for specific components

   • Combine configurations only when there are clear dependencies

   • Keep configurations reusable across different use cases

2. Clear Naming

   # Use descriptive names for nestd configs
   llm = hp.nest("configs/llm.py", name="llm")
   indexer = hp.nest("configs/indexer.py", name="indexer")

3. Value Dependencies

   # Explicitly pass dependent values
   retriever = hp.nest(
       "configs/retriever.py",
       values={"embedding_dim": embedder["embedding_dim"]}
   )

4. File Organization

   # Keep related configs in a dedicated directory
   configs/
   ├── llm.py
   ├── indexing.py
   ├── retrieval.py
   └── rag.py