Router DC (Dual-Contrastive)

Overview

router_dc is a semantic selection algorithm that matches queries to models using embedding similarity via dual-contrastive learning.

It aligns to config/algorithm/selection/router-dc.yaml.

Paper: Query-Based Router by Dual Contrastive Learning

Key Advantages

• Learns a semantic mapping between query embeddings and model capability embeddings.
• No explicit ranking rules needed — selection is driven by learned similarity.
• Supports both query-side and model-side contrastive learning.
• Useful when prompt semantics matter more than static priority or cost.

Algorithm Principle

Router-DC learns two embedding spaces — one for queries, one for models — and brings matching query-model pairs closer together using dual contrastive loss:

1. Query Embedding: Each user query is encoded into a dense vector via the configured embedding provider.
2. Model Embedding: Each model is represented by an embedding derived from its description and optional capability tags.
3. Contrastive Learning: Positive pairs (query, correct model) are pushed together; negative pairs (query, wrong model) are pushed apart.
4. Matching: At inference time, the query embedding is compared to all model embeddings using cosine similarity with temperature-scaled softmax:

$P(\text{model}_i \mid \text{query}) = \frac{\exp(\text{sim}(q, m_i) / \tau)}{\sum_j \exp(\text{sim}(q, m_j) / \tau)}$

Where $\tau$ is the temperature (temperature, default 0.07).

Select Flow

Model Embedding Initialization

Models need descriptions for embedding-based matching. Configure descriptions in modelCards:

routing:
  modelCards:
    - name: llama-3.2-1b
      description: "Fast small model for simple tasks, low cost"
      capabilities: ["summarization", "simple_qa"]
    - name: codellama-7b
      description: "Code generation specialist, good at programming tasks"
      capabilities: ["code_generation", "debugging"]

When use_capabilities: true, capability tags are concatenated with descriptions to enrich embeddings.

What Problem Does It Solve?

Some workloads are primarily semantic matching problems where the best model depends on the query meaning more than explicit heuristics. router_dc learns a query-to-model embedding space so routing follows semantic fit instead of only static priority or cost rules.

When to Use

• The best candidate depends on semantic similarity between prompt and model profile.
• You want a learned selector without full online exploration.
• One route should route by semantic fit rather than only cost or latency.
• Models have descriptive profiles or capability tags.

Known Limitations

• Requires model descriptions: If models lack descriptions, embedding quality degrades.
• Cold query problem: Rare query types may not match well with any model embedding.
• Affinity matrix: The query-model affinity matrix (affinityMatrix in code) is currently initialized but not actively updated online; it serves as a future extension point for online contrastive learning.
• Temperature sensitivity: Very low temperature makes the selector near-greedy; very high temperature makes it near-uniform.

Configuration

algorithm:
  type: router_dc
  router_dc:
    temperature: 0.07           # Softmax temperature (lower = sharper)
    dimension_size: 768         # Embedding dimension
    min_similarity: 0.3         # Minimum similarity threshold
    use_query_contrastive: true # Enable query-side contrastive learning
    use_model_contrastive: true # Enable model-side contrastive learning
    require_descriptions: false # Fail if models lack descriptions
    use_capabilities: true      # Include capability tags in embeddings

Parameters

Parameter	Type	Default	Description
temperature	float	0.07	Softmax temperature (lower = more confident selection)
dimension_size	int	768	Embedding dimension size
min_similarity	float	0.3	Minimum similarity threshold for valid matches (0–1)
use_query_contrastive	bool	true	Enable query-side contrastive learning
use_model_contrastive	bool	true	Enable model-side contrastive learning
require_descriptions	bool	false	Require all models to have descriptions
use_capabilities	bool	true	Include capability tags in embedding text

Feedback

Router-DC supports UpdateFeedback() for online affinity updates. When feedback arrives, the query-model affinity matrix is updated to reflect observed preferences:

curl -X POST http://localhost:8000/api/v1/feedback \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Write a Python function to sort a list",
    "winner_model": "codellama-7b",
    "decision_name": "coding"
  }'