LoRA Dispatcher

The dispatcher enables intelligent routing between multiple domain-specific adapters.

Architecture

Query → Sentence Encoder → Domain Router → Adapter Selection → Generation

Components

Sentence Encoder

Converts text to semantic embeddings:

Default: all-MiniLM-L6-v2
Can use any SentenceTransformer model
Embeddings cached for performance

Domain Router

Multi-class logistic regression classifier:

Trained on domain examples
Outputs confidence scores
Supports confidence thresholds

Domain Registry

Manages multiple adapters:

Domain name → adapter mapping
Example queries per domain
Metadata and configuration

Usage Pattern

1. Register Domains

dispatcher.register_domain(
    domain_name="medical",
    adapter_path="medical.air",
    examples=[
        "What are symptoms of flu?",
        "How does insulin work?",
    ]
)

2. Train Router

dispatcher.train_router()

3. Route Queries

domain, confidence, scores = dispatcher.route_query(
    "What is diabetes?"
)
# Returns: ("medical", 0.95, {"medical": 0.95, "legal": 0.05})

Confidence Thresholds

Set minimum confidence for domain selection:

dispatcher = LoRADispatcher(
    base_model=model,
    confidence_threshold=0.7
)

If confidence < threshold: * Returns None (no domain) * Can fallback to default adapter * Prevents incorrect routing

Embedding Cache

Cache embeddings for faster routing:

dispatcher = LoRADispatcher(
    base_model=model,
    cache_embeddings=True
)

Benefits: * Faster repeated queries * Reduced compute * Consistent routing

Domain Overlap Analysis

Analyze similarity between domains:

overlap = dispatcher.analyze_domain_overlap()
# Returns: {
#   ("medical", "legal"): 0.15,
#   ("medical", "tech"): 0.08,
#   ...
# }

Use to: * Identify ambiguous domains * Improve example selection * Optimize domain definitions