SVM (Support Vector Machine)

Overview

svm uses a trained linear or RBF support-vector classifier to map request features to a candidate model.

Implementation: Rust via Linfa (linfa-svm).

Key Advantages

• Learns explicit decision boundaries — interpretable via support vectors.
• RBF kernel captures non-linear patterns in query-to-model mapping.
• Lightweight inference compared to neural network approaches.
• Well-understood theoretical guarantees (maximum margin).

Algorithm Principle

SVM finds the hyperplane that maximizes the margin between different model classes:

$\min_{w, b} \frac{1}{2} \|w\|^2 + C \sum_{i} \xi_i$

$\text{s.t. } y_i(w^T \phi(x_i) + b) \geq 1 - \xi_i, \quad \xi_i \geq 0$

With the RBF (Radial Basis Function) kernel:

$K(x_i, x_j) = \exp(-\gamma \|x_i - x_j\|^2)$

The loaded RBF artifact contains the gamma used by its classifiers. Training also determines the support vectors, coefficients, and regularization.

Multi-class selection (more than 2 candidates) uses one-vs-rest classification.

Select Flow

What Problem Does It Solve?

Some workloads need a lightweight learned classifier with clearer decision boundaries than heuristic routing but less operational cost than deeper neural selectors. svm addresses that by learning margin-maximizing query-to-model boundaries over the routing features.

When to Use

• You have an SVM-based selector artifact for the route.
• Lightweight learned classification is enough for model choice.
• You want learned selection with interpretable decision boundaries.
• The query-to-model mapping has clear non-linear patterns.

Known Limitations

• Requires pre-training from historical query-to-model assignment data.
• RBF hyperparameters must be tuned while building the artifact; the Router does not retune them at request time.
• Multi-class SVM uses one-vs-rest, which can be suboptimal for many candidates.
• Does not support online learning — must be retrained for new patterns.

Configuration

algorithm:
  type: svm

Global ML Settings

global:
  router:
    model_selection:
      ml:
        models_path: ".cache/ml-models"
        embedding_dim: 768
        svm:
          kernel: rbf
          pretrained_path: .cache/ml-models/svm_model.json

Parameters

Parameter	Type	Default	Description
kernel	string	rbf	Empty-selector kernel: rbf (or gaussian) and linear are supported; other values fall back to linear
gamma	float	1.0	Accepted compatibility field; loading an artifact uses the gamma stored in that artifact rather than this value
pretrained_path	string	—	Path to pre-trained SVM model (JSON format)

Training

See ML Model Selection README for the training pipeline. SVM models are trained on labeled query-to-model assignment data using Linfa's SVM implementation.

Training examples and labels can contain sensitive request data; govern them and the derived artifact accordingly. See a complete example: config/fragments/algorithm/selection/svm.yaml.