LoRA Dispatcher

LoRA Dispatcher with Intelligent Routing

Automatically selects and applies the most suitable LoRA adapter based on query content using sentence embeddings and classification.

class ual_adapter.core.dispatcher.LoRADispatcher(encoder_model: str = 'all-MiniLM-L6-v2', confidence_threshold: float = 0.7, cache_embeddings: bool = True)

Bases: object

Intelligent dispatcher that routes queries to appropriate domain LoRAs.

Uses sentence embeddings and a trained classifier to determine which domain-specific adapter to use for a given query.

__init__(encoder_model: str = 'all-MiniLM-L6-v2', confidence_threshold: float = 0.7, cache_embeddings: bool = True)

Initialize the LoRA dispatcher.

Parameters:

• encoder_model – Sentence transformer model for embeddings

• confidence_threshold – Minimum confidence for domain selection

• cache_embeddings – Whether to cache computed embeddings

register_domain(domain_name: str, adapter_weights: Dict[str, Tensor], training_texts: List[str], metadata: Dict[str, Any] | None = None) → None

Register a domain-specific LoRA adapter.

Parameters:

• domain_name – Name of the domain

• adapter_weights – LoRA weights for this domain

• training_texts – Sample texts from this domain for routing

• metadata – Optional metadata about the domain

route_query(query: str, return_all_scores: bool = False) → Tuple[str | None, float, Dict[str, float] | None]

Determine which domain adapter to use for a query.

Parameters:

• query – The input query text

• return_all_scores – Whether to return scores for all domains

Returns:

Tuple of (selected_domain, confidence, all_scores_dict)

apply_adapter(model: Module, query: str, verbose: bool = True) → Tuple[Module, str, float]

Apply the appropriate adapter to a model based on the query.

Parameters:

• model – The base model to apply adapter to

• query – The input query

• verbose – Whether to log routing decisions

Returns:

Tuple of (model_with_adapter, selected_domain, confidence)

analyze_domain_overlap() → Dict[str, Any]

Analyze overlap between registered domains.

save(path: str) → None

Save dispatcher state to disk.

classmethod load(path: str) → LoRADispatcher

Load dispatcher state from disk.

class ual_adapter.core.dispatcher.DomainAdapter(name: str, weights: Dict[str, Tensor], metadata: Dict[str, Any])

Bases: object

Container for domain-specific adapter information.

__init__(name: str, weights: Dict[str, Tensor], metadata: Dict[str, Any])

Initialize domain adapter.

Parameters:

• name – Domain name

• weights – LoRA weights for this domain

• metadata – Additional information about the domain

get_info() → Dict[str, Any]

Get information about this domain adapter.

LoRADispatcher Class

class ual_adapter.core.dispatcher.LoRADispatcher(encoder_model: str = 'all-MiniLM-L6-v2', confidence_threshold: float = 0.7, cache_embeddings: bool = True)

Intelligent dispatcher that routes queries to appropriate domain LoRAs.

Uses sentence embeddings and a trained classifier to determine which domain-specific adapter to use for a given query.

__init__(encoder_model: str = 'all-MiniLM-L6-v2', confidence_threshold: float = 0.7, cache_embeddings: bool = True)

Initialize the LoRA dispatcher.

Parameters:

• encoder_model – Sentence transformer model for embeddings

• confidence_threshold – Minimum confidence for domain selection

• cache_embeddings – Whether to cache computed embeddings

register_domain(domain_name: str, adapter_weights: Dict[str, Tensor], training_texts: List[str], metadata: Dict[str, Any] | None = None) → None

Register a domain-specific LoRA adapter.

Parameters:

• domain_name – Name of the domain

• adapter_weights – LoRA weights for this domain

• training_texts – Sample texts from this domain for routing

• metadata – Optional metadata about the domain

route_query(query: str, return_all_scores: bool = False) → Tuple[str | None, float, Dict[str, float] | None]

Determine which domain adapter to use for a query.

Parameters:

• query – The input query text

• return_all_scores – Whether to return scores for all domains

Returns:

Tuple of (selected_domain, confidence, all_scores_dict)

apply_adapter(model: Module, query: str, verbose: bool = True) → Tuple[Module, str, float]

Apply the appropriate adapter to a model based on the query.

Parameters:

• model – The base model to apply adapter to

• query – The input query

• verbose – Whether to log routing decisions

Returns:

Tuple of (model_with_adapter, selected_domain, confidence)

analyze_domain_overlap() → Dict[str, Any]

Analyze overlap between registered domains.

save(path: str) → None

Save dispatcher state to disk.

classmethod load(path: str) → LoRADispatcher

Load dispatcher state from disk.

DomainAdapter Class

class ual_adapter.core.dispatcher.DomainAdapter(name: str, weights: Dict[str, Tensor], metadata: Dict[str, Any])

Container for domain-specific adapter information.

__init__(name: str, weights: Dict[str, Tensor], metadata: Dict[str, Any])

Initialize domain adapter.

Parameters:

• name – Domain name

• weights – LoRA weights for this domain

• metadata – Additional information about the domain

get_info() → Dict[str, Any]

Get information about this domain adapter.