What type of PR is this?

Feature enhancement for intelligent routing system

What this PR does / why we need it:

This PR implements intent-aware LoRA (Low-Rank Adaptation) routing support, enabling the semantic router to automatically select different LoRA adapters based on the classified intent/category of incoming requests.

Key Changes:

1. Configuration Structure:

   • Added LoRAAdapter struct to define available LoRA adapters per model in model_config
   • Added lora_name field to ModelScore for specifying which LoRA adapter to use
   • LoRA adapters must be pre-defined in model_config before being referenced

2. Validation:

   • Implemented validateLoRAName() to ensure lora_name references a valid LoRA defined in the model's configuration
   • Provides clear error messages with available LoRA names when validation fails

3. Model Selection Logic:

   • Updated selectBestModelInternal() to use LoRA name as the final model name when specified
   • Updated GetModelsForCategory() to return LoRA names for proper filtering
   • When a request is classified, the router now returns the LoRA adapter name instead of the base model name

4. Documentation & Examples:

   • Added comprehensive example configuration: config/intelligent-routing/in-tree/lora_routing_example.yaml
   • Updated documentation in website/docs/overview/categories/configuration.md
   • Updated README to reflect LoRA adapter routing capability

Configuration Example:

# Define LoRA adapters in model_config
model_config:
  "llama2-7b":
    reasoning_family: "llama2"
    loras:
      - name: "technical-lora"
        description: "Optimized for technical questions"
      - name: "medical-lora"
        description: "Specialized for medical domain"

# Reference them in categories
categories:
  - name: "technical"
    model_scores:
      - model: "llama2-7b"
        lora_name: "technical-lora"  # Routes to technical LoRA
        score: 1.0

How It Works:

1. LoRA adapters are defined in model_config under the base model
2. Request is classified into a category (e.g., "technical")
3. Router selects the best ModelScore for that category
4. Configuration validator ensures lora_name is defined in model's loras list
5. If lora_name is specified, it replaces the base model name
6. Request is sent to vLLM with model="technical-lora"
7. vLLM automatically routes to the appropriate LoRA adapter

Benefits:

• Domain Expertise: Fine-tuned adapters for specific domains (technical, medical, legal, etc.)
• Cost Efficiency: Share base model weights across adapters, reducing memory footprint
• Easy A/B Testing: Compare adapter versions by adjusting scores
• Flexible Deployment: Add/remove adapters without router restart
• Configuration Validation: Prevents typos and missing LoRA definitions

Prerequisites:

• vLLM server must be started with --enable-lora flag
• LoRA adapters must be registered using --lora-modules parameter
• LoRA names in config must match those registered with vLLM

Fixes: #545